eaddario/gemma-3-12b-it-pruned-GGUF
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

初始化项目，由ModelHub XC社区提供模型

Browse Source 
Model: eaddario/gemma-3-12b-it-pruned-GGUF
Source: Original Platform
This commit is contained in:
ModelHub XC
2026-07-21 18:50:10 +08:00
commit 66ebbdcd29
114 changed files with 24940 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
.gitattributes vendored Normal file
View File

@@ -0,0 +1,37 @@
*.7z filter=lfs diff=lfs merge=lfs -text
*.arrow filter=lfs diff=lfs merge=lfs -text
*.bin filter=lfs diff=lfs merge=lfs -text
*.bz2 filter=lfs diff=lfs merge=lfs -text
*.ckpt filter=lfs diff=lfs merge=lfs -text
*.ftz filter=lfs diff=lfs merge=lfs -text
*.gz filter=lfs diff=lfs merge=lfs -text
*.h5 filter=lfs diff=lfs merge=lfs -text
*.joblib filter=lfs diff=lfs merge=lfs -text
*.lfs.* filter=lfs diff=lfs merge=lfs -text
*.mlmodel filter=lfs diff=lfs merge=lfs -text
*.model filter=lfs diff=lfs merge=lfs -text
*.msgpack filter=lfs diff=lfs merge=lfs -text
*.npy filter=lfs diff=lfs merge=lfs -text
*.npz filter=lfs diff=lfs merge=lfs -text
*.onnx filter=lfs diff=lfs merge=lfs -text
*.ot filter=lfs diff=lfs merge=lfs -text
*.parquet filter=lfs diff=lfs merge=lfs -text
*.pb filter=lfs diff=lfs merge=lfs -text
*.pickle filter=lfs diff=lfs merge=lfs -text
*.pkl filter=lfs diff=lfs merge=lfs -text
*.pt filter=lfs diff=lfs merge=lfs -text
*.pth filter=lfs diff=lfs merge=lfs -text
*.rar filter=lfs diff=lfs merge=lfs -text
*.safetensors filter=lfs diff=lfs merge=lfs -text
*.gguf filter=lfs diff=lfs merge=lfs -text
*.logits filter=lfs diff=lfs merge=lfs -text
*.dat filter=lfs diff=lfs merge=lfs -text
*.tar.* filter=lfs diff=lfs merge=lfs -text
*.tar filter=lfs diff=lfs merge=lfs -text
*.tflite filter=lfs diff=lfs merge=lfs -text
*.tgz filter=lfs diff=lfs merge=lfs -text
*.wasm filter=lfs diff=lfs merge=lfs -text
*.xz filter=lfs diff=lfs merge=lfs -text
*.zip filter=lfs diff=lfs merge=lfs -text
*.zst filter=lfs diff=lfs merge=lfs -text
*tfevents* filter=lfs diff=lfs merge=lfs -text

584
.gitignore vendored Normal file
View File

@@ -0,0 +1,584 @@
# https://github.com/github/gitignore
# -------------
# General Linux
# -------------
.directory
.fuse_hidden*
.nfs*
.Trash-*
*~
# -----------
# General OSX
# -----------
._*
.apdisk
.AppleDB
.AppleDesktop
.AppleDouble
.com.apple.timemachine.donotpresent
.DocumentRevisions-V100
.DS_Store
.fseventsd
.LSOverride
.Spotlight-V100
.TemporaryItems
.Trashes
.VolumeIcon.icns
Network Trash Folder
Temporary Items
# ---------------
# General Windows
# ---------------
[Dd]esktop.ini
*.cab
*.lnk
*.msi
*.msix
*.msm
*.msp
*.stackdump
$RECYCLE.BIN/
ehthumbs_vista.db
ehthumbs.db
Thumbs.db
Thumbs.db:encryptable
# -----------------
# General JetBrains
# -----------------
.idea_modules/
.idea/
*.iml
*.ipr
*.iws
# ---------------
# General VS Code
# ---------------
!.vscode/*.code-snippets
!.vscode/extensions.json
!.vscode/launch.json
!.vscode/settings.json
!.vscode/tasks.json
.history/
.vscode/
*.vsix
# ---------------------
# General Visual Studio
# ---------------------
__pycache__/
_NCrunch_*
_pkginfo.txt
_Pvt_Extensions
_ReSharper*/
_TeamCity*
_UpgradeReport_Files/
!?*.[Cc]ache/
!.vscode/extensions.json
!**/[Pp]ackages/build/
.*crunch*.local.xml
.builds
.cr/personal
.fake/
.ionide/
.localhistory/
.mfractor/
.ntvs_analysis.dat
.paket/paket.exe
.sass-cache/
.vs/
.vscode/*
.vshistory/
[Aa][Rr][Mm]/
[Aa][Rr][Mm]64/
[Bb]in/
[Bb]uild[Ll]og.*
[Dd]ebug/
[Dd]ebugPS/
[Dd]ebugPublic/
[Ee]xpress/
[Ll]og/
[Ll]ogs/
[Oo]bj/
[Rr]elease/
[Rr]eleasePS/
[Rr]eleases/
[Tt]est[Rr]esult*/
[Ww][Ii][Nn]32/
*_h.h
*_i.c
*_p.c
*_wpftmp.csproj
*- [Bb]ackup ([0-9]).rdl
*- [Bb]ackup ([0-9][0-9]).rdl
*- [Bb]ackup.rdl
*.[Cc]ache
*.[Pp]ublish.xml
*.[Rr]e[Ss]harper
*.appx
*.appxbundle
*.appxupload
*.aps
*.azurePubxml
*.bim_*.settings
*.bim.layout
*.binlog
*.btm.cs
*.btp.cs
*.build.csdef
*.cachefile
*.code-workspace
*.coverage
*.coveragexml
*.dbmdl
*.dbproj.schemaview
*.dotCover
*.DotSettings.user
*.dsp
*.dsw
*.e2e
*.GhostDoc.xml
*.gpState
*.ilk
*.iobj
*.ipdb
*.jfm
*.jmconfig
*.ldf
*.mdf
*.meta
*.mm.*
*.ncb
*.ndf
*.nuget.props
*.nuget.targets
*.nupkg
*.nvuser
*.obj
*.odx.cs
*.opendb
*.opensdf
*.opt
*.pch
*.pdb
*.pfx
*.pgc
*.pgd
*.pidb
*.plg
*.psess
*.publishproj
*.publishsettings
*.pubxml
*.pyc
*.rdl.data
*.rptproj.bak
*.rptproj.rsuser
*.rsp
*.rsuser
*.sap
*.sbr
*.scc
*.sdf
*.sln.docstates
*.sln.iml
*.snupkg
*.suo
*.svclog
*.tlb
*.tlh
*.tli
*.tlog
*.tmp
*.tmp_proj
*.tss
*.user
*.userosscache
*.userprefs
*.vbp
*.vbw
*.VC.db
*.VC.VC.opendb
*.VisualState.xml
*.vsp
*.vspscc
*.vspx
*.vssscc
*.xsd.cs
**/[Pp]ackages/*
**/*.DesktopClient/GeneratedArtifacts
**/*.DesktopClient/ModelManifest.xml
**/*.HTMLClient/GeneratedArtifacts
**/*.Server/GeneratedArtifacts
**/*.Server/ModelManifest.xml
~$*
$tf/
AppPackages/
artifacts/
ASALocalRun/
AutoTest.Net/
Backup*/
BenchmarkDotNet.Artifacts/
bld/
BundleArtifacts/
ClientBin/
coverage*.info
coverage*.json
coverage*.xml
csx/
dlldata.c
DocProject/buildhelp/
DocProject/Help/*.hhc
DocProject/Help/*.hhk
DocProject/Help/*.hhp
DocProject/Help/*.HxC
DocProject/Help/*.HxT
DocProject/Help/html
DocProject/Help/Html2
ecf/
FakesAssemblies/
FodyWeavers.xsd
Generated_Code/
healthchecksdb
ipch/
MigrationBackup/
mono_crash.*
nCrunchTemp_*
node_modules/
nunit-*.xml
OpenCover/
orleans.codegen.cs
Package.StoreAssociation.xml
paket-files/
project.fragment.lock.json
project.lock.json
publish/
PublishScripts/
rcf/
ScaffoldingReadMe.txt
ServiceFabricBackup/
StyleCopReport.xml
TestResult.xml
UpgradeLog*.htm
UpgradeLog*.XML
x64/
x86/
# ----------------------
# General Archived Files
# ----------------------
*.7z
*.bz2
*.bzip
*.bzip2
*.deb
*.dmg
*.egg
*.gem
*.gz
*.gzip
*.iso
*.jar
*.lzma
*.rar
*.rpm
*.tar
*.tgz
*.txz
*.tzst
*.xar
*.xpi
*.xz
*.zip
*.zst
# -----
# C/C++
# -----
.tmp_versions/
*.a
*.app
*.cmd
*.d
*.dll
*.dSYM/
*.dylib
*.elf
*.exe
*.exp
*.gch
*.hex
*.i*86
*.idb
*.ko
*.la
*.lai
*.lib
*.lo
*.map
*.mod*
*.o
*.out
*.slo
*.so
*.so.*
*.su
*.x86_64
dkms.conf
Mkfile.old
Module.symvers
modules.order
# ----
# CUDA
# ----
*.cubin
*.fatbin
*.gpu
*.i
*.ii
*.ptx
# --
# Go
# --
.env
*.exe~
*.test
go.work
go.work.sum
# ----
# Java
# ----
*.class
*.ctxt
*.ear
*.hprof
*.nar
*.tar.gz
*.war
hs_err_pid*
replay_pid*
# -----
# Julia
# -----
*.jl.*.cov
*.jl.cov
*.jl.mem
deps/build.log
deps/deps.jl
deps/downloads/
deps/src/
deps/usr/
docs/build/
docs/site/
Manifest.toml
# ------------------------------
# JavaScript / Node / TypeScript
# ------------------------------
.cache
.cache/
.docusaurus
.dynamodb/
.env.development.local
.env.local
.env.production.local
.env.test.local
.eslintcache
.fusebox/
.grunt
.lock-wscript
.next
.node_repl_history
.npm
.nuxt
.nyc_output
.parcel-cache
.pnp.*
.pnpm-debug.log*
.rpt2_cache/
.rts2_cache_cjs/
.rts2_cache_es/
.rts2_cache_umd/
.serverless/
.stylelintcache
.temp
.tern-port
.vscode-test
.vuepress/dist
.yarn-integrity
.yarn/build-state.yml
.yarn/cache
.yarn/install-state.gz
.yarn/unplugged
*.lcov
*.pid
*.pid.lock
*.seed
*.tsbuildinfo
bower_components
dist
jspm_packages/
lerna-debug.log*
lib-cov
logs
npm-debug.log*
out
pids
report.[0-9]*.[0-9]*.[0-9]*.[0-9]*.json
web_modules/
yarn-debug.log*
yarn-error.log*
# ------
# Python
# ------
__pypackages__/
.coverage
.coverage.*
.dmypy.json
.eggs/
.hypothesis/
.installed.cfg
.ipynb_checkpoints
.mypy_cache/
.nox/
.pdm-build/
.pdm-python
.pdm.toml
.pybuilder/
.pyre/
.pytest_cache/
.Python
.python-version
.pytype/
.ropeproject
.scrapy
.spyderproject
.spyproject
.tox/
.venv
.webassets-cache
*.cover
*.egg-info/
*.manifest
*.mo
*.pot
*.py,cover
*.py[cod]
*.sage.py
*.spec
*$py.class
/site
build/
celerybeat-schedule
celerybeat.pid
cover/
coverage.xml
cython_debug/
db.sqlite3
db.sqlite3-journal
develop-eggs/
dist/
dmypy.json
docs/_build/
downloads/
eggs/
env.bak/
env/
htmlcov/
instance/
ipython_config.py
lib/
lib64/
local_settings.py
MANIFEST
nosetests.xml
parts/
pip-delete-this-directory.txt
pip-log.txt
profile_default/
sdist/
share/python-wheels/
target/
var/
venv.bak/
venv/
wheels/
# ----
# Rust
# ----
**/*.rs.bk
Cargo.lock
debug/
# -----
# Scala
# -----
# -----
# CMake
# -----
_deps
cmake_install.cmake
CMakeCache.txt
CMakeFiles
CMakeLists.txt.user
CMakeScripts
CMakeUserPresets.json
compile_commands.json
CTestTestfile.cmake
install_manifest.txt
Makefile
Testing
# ------
# Gradle
# ------
!gradle-wrapper.jar
!gradle-wrapper.properties
!src/**/build/
.classpath
.gradle
.gradletasknamecache
.project
**/build/
gradle-app.setting
# -----
# Maven
# -----
.mvn/
buildNumber.properties
dependency-reduced-pom.xml
pom.xml.next
pom.xml.releaseBackup
pom.xml.tag
pom.xml.versionsBackup
release.properties
# ---------
# Terraform
# ---------
.terraform.tfstate.lock.info
.terraform/
.terraformrc
*_override.tf
*_override.tf.json
*.tfstate
*.tfstate.*
*.tfvars
*.tfvars.json
crash.*.log
crash.log
override.tf
override.tf.json
terraform.rc

229
README.md Normal file
View File

@@ -0,0 +1,229 @@
---
base_model:
- google/gemma-3-12b-it
datasets:
- eaddario/imatrix-calibration
language:
- en
license:
- gemma
pipeline_tag: text-generation
tags:
- gguf
- quant
- pruned
- experimental
---
# Experimental layer-wise + pruned (layers 26 and 29) quantization of google/gemma-3-12b-it
Using [LLaMA C++][llm] release [b5540][llm-rel] for quantization.
Original model: [google/gemma-3-12b-it][mdl]
From the original model creators:
> Terms of Use: [Terms](https://ai.google.dev/gemma/terms)
>
> Authors: Google DeepMind
>
> Model Information
> Summary description and brief definition of inputs and outputs.
>
> Description
> Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. Gemma 3 models are multimodal, handling text and image input and generating text output, with open weights for both pre-trained variants and instruction-tuned variants. Gemma 3 has a large, 128K context window, multilingual support in over 140 languages, and is available in more sizes than previous versions. Gemma 3 models are well-suited for a variety of text generation and image understanding tasks, including question answering, summarization, and reasoning. Their relatively small size makes it possible to deploy them in environments with limited resources such as laptops, desktops or your own cloud infrastructure, democratizing access to state of the art AI models and helping foster innovation for everyone.
>
> Inputs and outputs
> Input:
>
> Text string, such as a question, a prompt, or a document to be summarized
> Images, normalized to 896 x 896 resolution and encoded to 256 tokens each
> Total input context of 128K tokens for the 4B, 12B, and 27B sizes, and 32K tokens for the 1B size
> Output:
>
> Generated text in response to the input, such as an answer to a question, analysis of image content, or a summary of a document
> Total output context of 8192 tokens
Usage
# PLEASE READ THIS BEFORE USING THESE EXPERIMENTAL VERSIONS!
An area of personal interest is finding ways to optimize the inference performance of LLMs when deployed in resource-constrained environments like commodity hardware, desktops, laptops, mobiles, edge devices, etc. There are many approaches to accomplish this, including architecture simplification and knowledge distillation, but my focus has been primarily on quantization and pruning.
The method used to produce these experimental versions is covered in [Squeezing Tensor Bits: the quest for smaller LLMs][mdm], but at a high level it involves using a custom version of `llama-imatrix` and `llama-quantize` to identify influential tensors, quantize the most important layers to higher bit precision and the less important to lower bits, and remove (prune) one or more layers. This process was partly inspired by Dumitru's et al [Layer-Wise Quantization: A Pragmatic and Effective Method for Quantizing LLMs Beyond Integer Bit-Levels][lwq-ppr], and Xin Men's et al [ShortGPT: Layers in Large Language Models are More Redundant Than You Expect][sgpt-ppr]
As of version [b5125][qtz-rel], [llama-quantize][qtz] can now perform **tensor-wide quantization (TWQ)**, whereby user-defined tensors are quantized at a specific level, or perform **layer-wise quantization (LWQ)** by selecting different quantization types per tensor/layer. For example, `--tensor-type attn_v=q6_k` will quantize all *Attention Value* tensors at *q6_k* (TWQ), and `--tensor-type "\.([0-9]|1[01257]|31)\.attn_k=q4_k"` will quantize *Attention Key* tensors on layers 0 to 9, 10, 11, 12, 15, 17 and 31 at *q4_k*, leaving the remaining layers at their default value (LWQ).
A custom version of `llama-quantize` is used to prune the model by providing a comma-separated list in the `--prune-layers` command line option. The pruning operation will renumber remaining layers to avoid gaps in the sequence, update the relevant model metadata and, if an imatrix is available, it will use the correct importance score vector. This option can be used alongside `--tensor-type` to perform tensor/layer-wise quantization on selected tensor types, whilst at the same time pruning others. For example:
```
llama-quantize --tensor-type attn=q6_k --prune-layers 3,7,11 --imatrix imatrix.dat model-f32.gguf model-q4_k_m.gguf q4_k_m
```
An enhanced version of [llama-imatrix][imx] generates useful statistics to guide the tensor and layer selection process. `--show-statistics` will display:
### Tensor statistics:
- **Σ(Bias):** the sum of all activations over the tensor (i.e. the Importance Scores)
- **Min & Max:** minimum and maximum activation values
- **μ & σ:** activations' mean and standard deviation
- **% Active:** proportion of elements whose average activation exceeds a very small threshold (1e-6). Helpful to determine how alive/dormant the tensor is during inference
- **N:** number of activations in the tensor
- **Entropy:** entropy of the activation distribution, in bits (standard Shannon entropy measurement)
- **E (norm):** Normalized entropy.
- **ZD Score:** z-score distribution as described in [*3.1 Layer Importance Scores*][lwq-lim] in the **Layer-Wise Quantization** paper
- **CosSim:** cosine similarity between same type tensors with respect to the previous layer (i.e. blk.7.attn_k and blk.6.attn_k)
### Layer statistics:
- **Σ(Bias):** weighted average of the sum of all activations over the tensor (i.e. the Importance Scores)
- **ZD Score:** weighted average of the z-score distribution as described in [*3.1 Layer Importance Scores*][lwq-lim] in the **Layer-Wise Quantization** paper
- **CosSim:** weighted average of the cosine similarity of whole layer with respect to the previous layer (i.e. Layer 3 and 2)
Please note that tensor statistics are calculated for each individual tensor and should be used to compare between tensors of the same type only. For example, assuming that *attn_k* in layer 10 has a higher influence during inference than *attn_k* in layer 7 because its **Σ(Bias)** is larger makes sense, whilst concluding the same between *attn_k* and *ffn_down* does not.
There are two Pull Request ([prune][prn-pr] and [imatrix][imtx-pr]) to merge these changes back into the core llama.cpp project. This may or may not ever happen so, until then, the modified version will be available on [GitHub][gh] ([prune][gh-prn] and [imatrix][gh-itx]).
For testing and comparison I use models produced by [Unsloth][ust] ([Daniel and Michael Han][ust-ai] do some really advanced level stuff!) and [Bartowski][btk] (see credits below) but when they don't provide versions of the required model, all tests and comparisons are done against naive quantizations obtained by simply running `llama-quantize` with no further optimization. In this case however, whilst both have versions of this model, Unsloth's uses a different vocabulary size on their quants (262144 vs 262208) which does not allow for a valid like-for-like comparison.
All experimental versions were generated using an appropriate imatrix created from calibration datasets available at [eaddario/imatrix-calibration][ical]. At its core, an Importance Matrix (imatrix) is a table or, more broadly, a structured representation that scores the relative importance of different features or parameters in a machine learning model. It essentially quantifies the "impact" each feature has on a specific outcome, prediction, or relationship being modelled, and it helps to counterbalance the negative effects of quantization and pruning.
The process to generate these models is roughly as follows:
1. Convert the original model's tensors to [GGUF][ggf] F16*
2. Estimate the Perplexity score for the F16 model (baseline) using the [wikitext-2-raw-v1][wki-dat] dataset, and save the [logits][lgt]
3. Generate an [imatrix][imx-dat] from selected calibration datasets
4. Determine tensor and layer Importance Score contribution using the modified version of `llama-imatrix`
5. Select an appropriate quant level for each tensor and quantize/prune the model using `llama-quantize`. In this model's case, layers **26** and **29** have been pruned
6. Calculate Perplexity, KL Divergence, ARC (Easy+Challenge), HellaSwag, MMLU, Truthful QA and WinoGrande scores for each quantized model
7. Keep versions with the best scores
8. Repeat until all desired quants are created. I find that quantizations below Q3/IQ3 are not fit for my purposes and therefore do not usually generate them, but happy to provide other quants on request.
*[BF16][bf16] would be preferred, but Apple's GPUs don't support it yet, and therefore any operations are executed in the CPU, making it unacceptably slow. This is expected to change in the near term but until then, if you are using Apple kit avoid using any models tagged BF16
# Models
### Sizes (in GB)
| Model | Bartowski | Repo | Shrinkage |
| ------------------------------------------------------------------- | --------: | ----: | --------: |
| [gemma-3-12b-it-pruned-IQ3_M](./gemma-3-12b-it-pruned-IQ3_M.gguf) | 5.66 | 5.21 | 8.0% |
| [gemma-3-12b-it-pruned-IQ3_S](./gemma-3-12b-it-pruned-IQ3_S.gguf) | 5.21 | 5.04 | 3.3% |
| [gemma-3-12b-it-pruned-IQ4_NL](./gemma-3-12b-it-pruned-IQ4_NL.gguf) | 6.89 | 6.16 | 10.6% |
| [gemma-3-12b-it-pruned-Q3_K_L](./gemma-3-12b-it-pruned-Q3_K_L.gguf) | 6.48 | 5.33 | 17.7% |
| [gemma-3-12b-it-pruned-Q3_K_M](./gemma-3-12b-it-pruned-Q3_K_M.gguf) | 6.01 | 5.04 | 16.1% |
| [gemma-3-12b-it-pruned-Q3_K_S](./gemma-3-12b-it-pruned-Q3_K_S.gguf) | 5.46 | 4.81 | 11.9% |
| [gemma-3-12b-it-pruned-Q4_K_M](./gemma-3-12b-it-pruned-Q4_K_M.gguf) | 7.30 | 6.20 | 15.1% |
| [gemma-3-12b-it-pruned-Q4_K_S](./gemma-3-12b-it-pruned-Q4_K_S.gguf) | 6.94 | 6.17 | 11.1% |
| [gemma-3-12b-it-pruned-Q5_K_M](./gemma-3-12b-it-pruned-Q5_K_M.gguf) | 8.44 | 7.32 | 13.3% |
| [gemma-3-12b-it-pruned-Q5_K_S](./gemma-3-12b-it-pruned-Q5_K_S.gguf) | 8.23 | 7.26 | 11.8% |
| [gemma-3-12b-it-pruned-Q6_K](./gemma-3-12b-it-pruned-Q6_K.gguf) | 9.66 | 9.01 | 6.7% |
| [gemma-3-12b-it-pruned-Q8_0](./gemma-3-12b-it-pruned-Q8_0.gguf) | 12.50 | 10.98 | 12.2% |
### Perplexity and KL Divergence scores
| Model | μPPL | 𝜌PPL | μKLD | RMS Δp |
| ------------------------------------------------------------------- | ------------------: | -----: | -----------------: | ------------: |
| [gemma-3-12b-it-pruned-IQ3_M](./gemma-3-12b-it-pruned-IQ3_M.gguf) | 35.478449 ±0.411747 | 80.18% | 1.229937 ±0.004277 | 30.166 ±0.089 |
| [gemma-3-12b-it-pruned-IQ3_S](./gemma-3-12b-it-pruned-IQ3_S.gguf) | 32.557804 ±0.369351 | 80.53% | 1.196823 ±0.004146 | 29.810 ±0.088 |
| [gemma-3-12b-it-pruned-IQ4_NL](./gemma-3-12b-it-pruned-IQ4_NL.gguf) | 36.879518 ±0.436763 | 80.79% | 1.190922 ±0.004215 | 29.578 ±0.088 |
| [gemma-3-12b-it-pruned-Q3_K_L](./gemma-3-12b-it-pruned-Q3_K_L.gguf) | 33.537202 ±0.371564 | 79.10% | 1.263077 ±0.004253 | 31.084 ±0.088 |
| [gemma-3-12b-it-pruned-Q3_K_M](./gemma-3-12b-it-pruned-Q3_K_M.gguf) | 34.029948 ±0.376082 | 78.62% | 1.300974 ±0.004321 | 31.537 ±0.088 |
| [gemma-3-12b-it-pruned-Q3_K_S](./gemma-3-12b-it-pruned-Q3_K_S.gguf) | 36.103787 ±0.402250 | 77.13% | 1.395578 ±0.004559 | 32.089 ±0.088 |
| [gemma-3-12b-it-pruned-Q4_K_M](./gemma-3-12b-it-pruned-Q4_K_M.gguf) | 31.219358 ±0.342580 | 80.83% | 1.106608 ±0.003932 | 29.739 ±0.088 |
| [gemma-3-12b-it-pruned-Q4_K_M-bartowski][b-q4km] | 9.197963 ±0.073909 | 98.63% | 0.045560 ±0.000416 | 6.400 ±0.054 |
| [gemma-3-12b-it-pruned-Q4_K_S](./gemma-3-12b-it-pruned-Q4_K_S.gguf) | 31.245217 ±0.343080 | 80.86% | 1.105266 ±0.003926 | 29.698 ±0.088 |
| [gemma-3-12b-it-pruned-Q5_K_M](./gemma-3-12b-it-pruned-Q5_K_M.gguf) | 30.668770 ±0.335889 | 81.60% | 1.039101 ±0.003848 | 29.177 ±0.088 |
| [gemma-3-12b-it-pruned-Q5_K_S](./gemma-3-12b-it-pruned-Q5_K_S.gguf) | 30.644285 ±0.335716 | 81.62% | 1.038541 ±0.003836 | 29.119 ±0.088 |
| [gemma-3-12b-it-pruned-Q6_K](./gemma-3-12b-it-pruned-Q6_K.gguf) | 30.107026 ±0.329175 | 81.96% | 1.009233 ±0.003785 | 28.800 ±0.088 |
| [gemma-3-12b-it-pruned-Q8_0](./gemma-3-12b-it-pruned-Q8_0.gguf) | 30.014168 ±0.327772 | 82.04% | 1.004325 ±0.003776 | 28.772 ±0.088 |
| [gemma-3-12b-it-pruned-F16](./gemma-3-12b-it-pruned-F16.gguf) | 9.042786 ±0.071503 | 100% | N/A | N/A |
### ARC, HellaSwag, MMLU, Truthful QA and WinoGrande scores
Scores generated using [llama-perplexity][ppl] with 750 tasks per test, and a context size of 768 tokens.
For the test data used in the generation of these scores, follow the appropriate links: [HellaSwag][hsw-tst], [ARC, MMLU, Truthful QA][tst-dat] and [WinoGrande][wng-tst]
| Model | ARC | HellaSwag | MMLU | Truthful QA | WinoGrande | Avg Score |
| ------------------------------------------------------------------- | --------------: | --------: | --------------: | --------------: | --------------: | --------: |
| [gemma-3-12b-it-pruned-IQ3_M](./gemma-3-12b-it-pruned-IQ3_M.gguf) | 66.9333 ±1.7190 | 74.67 | 42.0000 ±1.8034 | 39.4667 ±1.7860 | 67.4667 ±1.7119 | 58.11 |
| [gemma-3-12b-it-pruned-IQ3_S](./gemma-3-12b-it-pruned-IQ3_S.gguf) | 65.0667 ±1.7420 | 74.80 | 42.8000 ±1.8079 | 40.4000 ±1.7930 | 66.1333 ±1.7292 | 57.84 |
| [gemma-3-12b-it-pruned-IQ4_NL](./gemma-3-12b-it-pruned-IQ4_NL.gguf) | 68.2667 ±1.7007 | 75.47 | 42.6667 ±1.8072 | 40.1333 ±1.7910 | 68.5333 ±1.6968 | 59.01 |
| [gemma-3-12b-it-pruned-Q3_K_L](./gemma-3-12b-it-pruned-Q3_K_L.gguf) | 67.8667 ±1.7063 | 75.73 | 42.5333 ±1.8065 | 40.2667 ±1.7920 | 67.0667 ±1.7172 | 58.69 |
| [gemma-3-12b-it-pruned-Q3_K_M](./gemma-3-12b-it-pruned-Q3_K_M.gguf) | 68.6667 ±1.6949 | 75.20 | 44.4000 ±1.8155 | 40.0000 ±1.7900 | 65.8667 ±1.7325 | 58.83 |
| [gemma-3-12b-it-pruned-Q3_K_S](./gemma-3-12b-it-pruned-Q3_K_S.gguf) | 67.6000 ±1.7100 | 75.20 | 43.7333 ±1.8126 | 39.8667 ±1.7890 | 66.5333 ±1.7242 | 58.59 |
| [gemma-3-12b-it-pruned-Q4_K_M](./gemma-3-12b-it-pruned-Q4_K_M.gguf) | 69.8667 ±1.6766 | 77.47 | 43.6000 ±1.8119 | 40.8000 ±1.7958 | 68.4000 ±1.6988 | 60.03 |
| [gemma-3-12b-it-pruned-Q4_K_M-bartowski][b-q4km] | 69.6000 ±1.6807 | 81.07 | 43.6000 ±1.8119 | 41.0667 ±1.7976 | 75.7333 ±1.5664 | 62.21 |
| [gemma-3-12b-it-pruned-Q4_K_S](./gemma-3-12b-it-pruned-Q4_K_S.gguf) | 69.8667 ±1.6766 | 77.33 | 43.4667 ±1.8113 | 40.8000 ±1.7958 | 69.6000 ±1.6807 | 60.21 |
| [gemma-3-12b-it-pruned-Q5_K_M](./gemma-3-12b-it-pruned-Q5_K_M.gguf) | 68.8000 ±1.6929 | 77.87 | 45.3333 ±1.8190 | 41.2000 ±1.7984 | 68.8000 ±1.6929 | 60.40 |
| [gemma-3-12b-it-pruned-Q5_K_S](./gemma-3-12b-it-pruned-Q5_K_S.gguf) | 68.9333 ±1.6909 | 77.73 | 45.0667 ±1.8180 | 41.3333 ±1.7993 | 68.6667 ±1.6949 | 60.35 |
| [gemma-3-12b-it-pruned-Q6_K](./gemma-3-12b-it-pruned-Q6_K.gguf) | 68.9333 ±1.6909 | 78.66 | 45.4667 ±1.8194 | 40.8000 ±1.7958 | 68.6667 ±1.6949 | 60.51 |
| [gemma-3-12b-it-pruned-Q8_0](./gemma-3-12b-it-pruned-Q8_0.gguf) | 68.8000 ±1.6929 | 78.40 | 45.4667 ±1.8194 | 41.7333 ±1.8018 | 68.8000 ±1.6929 | 60.64 |
| [gemma-3-12b-it-pruned-F16](./gemma-3-12b-it-pruned-F16.gguf) | 69.2000 ±1.6869 | 81.20 | 45.3333 ±1.8190 | 41.4667 ±1.8002 | 74.8000 ±1.5864 | 62.40 |
### Tokens per Second - Benchmarks
Scores generated using [llama-bench][bch]. Naive (`llama-quantize` with no optimization) Q4_K_M quantization included for comparison.
| model | size | params | backend | threads | test | t/s |
| ------------------------------------------------------------------- | -------: | ------: | ---------- | ------: | ------------: | ------------: |
| [gemma-3-12b-it-pruned-Q4_K_M](./gemma-3-12b-it-pruned-Q4_K_M.gguf) | 5.77 GiB | 11.32 B | Metal,BLAS | 12 | pp512 | 523.27 ±3.74 |
| [gemma-3-12b-it-pruned-Q4_K_M](./gemma-3-12b-it-pruned-Q4_K_M.gguf) | 5.77 GiB | 11.32 B | Metal,BLAS | 12 | tg128 | 46.92 ±0.60 |
| [gemma-3-12b-it-pruned-Q4_K_M](./gemma-3-12b-it-pruned-Q4_K_M.gguf) | 5.77 GiB | 11.32 B | Metal,BLAS | 12 | pp1024+tg1024 | 75.45 ±0.35 |
| [gemma-3-12b-it-pruned-Q4_K_M-bartowski][b-q4km] | 6.79 GiB | 11.77 B | Metal,BLAS | 12 | pp512 | 482.36 ±23.04 |
| [gemma-3-12b-it-pruned-Q4_K_M-bartowski][b-q4km] | 6.79 GiB | 11.77 B | Metal,BLAS | 12 | tg128 | 45.54 ±1.21 |
| [gemma-3-12b-it-pruned-Q4_K_M-bartowski][b-q4km] | 6.79 GiB | 11.77 B | Metal,BLAS | 12 | pp1024+tg1024 | 74.27 ±0.19 |
# Metrics used
**[Perplexity][ppx]:** one of the key metrics used in NLP evaluation. It measures the quality of a language model by evaluating how well it predicts the next token given a particular sequence of words. A PPL of **1** indicates an exact match between predicted and actual, whereas values greater than one indicate a degree of "surprise" the generated token differs from the expected.
**[Kullback–Leibler (KL) Divergence][kld]:** a statistical measure of how much a probability distribution differs from another. When quantizing models (or altering the original tensors in any way for that matter), the closest we can preserve the weights' probability distribution to the original model the better, thus the closest to **0** the better.
**[AI2 Reasoning Challenge (ARC)][arc]:** a benchmark to evaluate the ability of AI models to answer complex science questions that require logical reasoning beyond pattern matching.
**[HellaSwag][hsw]:** the Harder Endings, Longer contexts, and Low-shot Activities for Situations With Adversarial Generations (bit of a mouthful!) is a benchmark designed to test commonsense natural language inference. It requires the model to predict the most likely ending of a sentence.
**[MMLU][mmlu]:** the Massive Multitask Language Understanding evaluates LLMs’ general knowledge and problem-solving abilities across 57 subjects, including elementary mathematics, US history, computer science, and law.
**[Truthful QA][tqa]:** evaluates how well LLMs generate truthful responses to questions. It identifies whether AI models can avoid generating false or misleading information, particularly in areas where human knowledge is prone to misconceptions.
**[Winogrande][wng]:** based on the [Winograd Schema Challenge][wng-chl], is a natural language understanding task requiring models to resolve ambiguities in sentences involving pronoun references.
## Credits
[LLaMa C++][llm] has a large and vibrant community of [contributors][llm-ctt] (~1,200 last time I checked) that actively maintain and extend its functionality, adding new models and architectures almost as fast as they appear (considering the breakneck speed at which the AI/ML field is advancing, this alone is a remarkable feat!), and whilst I'm grateful to each and everyone of them, I want to recognise three people in particular: **Thank You!** [Colin Kealty][btk] for the many contributions and for being one of the best sources of high quality quantized models available on Hugging Face, and a really big ***Thank You!*** to [Georgi Gerganov][ggg] for his amazing work with **llama.cpp** and the **ggml/gguf** libraries, and [Iwan Kawrakow][ikk] for being one of the key authors behind the many quantisation algorithms and the imatrix functionality.
[arc]: https://leaderboard.allenai.org/arc/submissions/get-started
[btk]: https://huggingface.co/bartowski
[bch]: https://github.com/ggml-org/llama.cpp/tree/master/tools/llama-bench
[bf16]: https://en.wikipedia.org/wiki/Bfloat16_floating-point_format
[b-q4km]: https://huggingface.co/bartowski/google_gemma-3-12b-it-pruned-GGUF/blob/main/google_gemma-3-12b-it-pruned-Q4_K_M.gguf
[u-q4km]: https://huggingface.co/unsloth
[ical]: https://huggingface.co/datasets/eaddario/imatrix-calibration
[ikk]: https://github.com/ikawrakow
[ggg]: https://github.com/ggerganov
[ggf]: https://huggingface.co/docs/hub/en/gguf
[gh]: https://github.com/EAddario
[gh-itx]: https://github.com/EAddario/llama.cpp/tree/imatrix
[gh-prn]: https://github.com/EAddario/llama.cpp/tree/prune
[hsw]: https://rowanzellers.com/hellaswag
[hsw-tst]: https://github.com/klosax/hellaswag_text_data
[imx-dat]: https://huggingface.co/eaddario/gemma-3-12b-it-pruned-GGUF/tree/main/imatrix
[imx]: https://github.com/ggml-org/llama.cpp/tree/master/tools/imatrix
[imtx-pr]: https://github.com/ggml-org/llama.cpp/pull/12718
[kld]: https://en.wikipedia.org/wiki/Kullback–Leibler_divergence
[llm]: https://github.com/ggerganov/llama.cpp
[llm-rel]: https://github.com/ggml-org/llama.cpp/releases/tag/b5540
[lgt]: https://huggingface.co/eaddario/gemma-3-12b-it-pruned-GGUF/tree/main/logits
[llm-ctt]: https://github.com/ggml-org/llama.cpp/graphs/contributors
[lwq-lim]: https://arxiv.org/html/2403.03853v3#S3
[lwq-ppr]: https://arxiv.org/abs/2406.17415
[mdm]: https://medium.com/@eaddario/squeezing-tensor-bits-the-quest-for-smaller-llms-86b23bd052ca
[mmlu]: https://github.com/hendrycks/test
[mdl]: https://huggingface.co/google/gemma-3-12b-it
[ppl]: https://github.com/ggml-org/llama.cpp/tree/master/tools/perplexity
[ppx]: https://huggingface.co/docs/transformers/en/perplexity
[prn-pr]: https://github.com/ggml-org/llama.cpp/pull/13037
[qtz]: https://github.com/ggml-org/llama.cpp/tree/master/tools/quantize
[qtz-rel]: https://github.com/ggerganov/llama.cpp/releases/tag/b5125
[sgpt-ppr]: https://arxiv.org/abs/2403.03853
[tst-dat]: https://huggingface.co/datasets/ikawrakow/validation-datasets-for-llama.cpp/tree/main
[tqa]: https://github.com/sylinrl/TruthfulQA
[ust]: https://huggingface.co/unsloth
[ust-ai]: https://unsloth.ai
[wng-chl]: https://cdn.aaai.org/ocs/4492/4492-21843-1-PB.pdf
[wki-dat]: https://huggingface.co/datasets/Salesforce/wikitext/tree/main/wikitext-2-raw-v1
[wng]: https://winogrande.allenai.org
[wng-tst]: https://huggingface.co/datasets/ikawrakow/winogrande-eval-for-llama.cpp/tree/main

3
gemma-3-12b-it-F16.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1ea2f961c2204bab46c86b8e34dfb3c5216808ab17b0434a0535cfb316197930
size 23539658496

3
gemma-3-12b-it-pruned-IQ3_M.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:352481a5f4781c56fa0a17f263d19e64633fb6ff9b73585d31f3f8569ee0642a
size 5207677472

3
gemma-3-12b-it-pruned-IQ3_S.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:6ee25db0fce1562dfa50126a76cf1bc6729f86a02dbad905c254a68be014e060
size 5039331872

3
gemma-3-12b-it-pruned-IQ4_NL.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:100f615fcbab82cae935425f69757eaf1d61508fd27e62514067b4504bda982c
size 6161594912

3
gemma-3-12b-it-pruned-Q3_K_L.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:4d02c8159953ceb36a3903622cb661ace7668124c9e0e6352a654bd32768d428
size 5327731232

3
gemma-3-12b-it-pruned-Q3_K_M.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:677b204f96f60a67204d3061920cec9b0a6a9e72d6abd03d46b766b0d363c255
size 5037365792

3
gemma-3-12b-it-pruned-Q3_K_S.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e2f944b09f0df99cc3d267a2653450e26ac32150b62380bfead1dd3feec7fce7
size 4813969952

3
gemma-3-12b-it-pruned-Q4_K_M.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:cabc37cfe948bb8e8c42673c152b0c5a010a8c97086ca9a5459b6cd262778a53
size 6204357152

3
gemma-3-12b-it-pruned-Q4_K_S.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:33cc8aee36a24491ce3570a83abf9df6ec0f6d016b94dfc731e6f77765fd9d91
size 6167001632

3
gemma-3-12b-it-pruned-Q5_K_M.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c197fcafbc748dfa8962a565b969a8f136a376657118c0eaf4142e88eaf41f11
size 7320291872

3
gemma-3-12b-it-pruned-Q5_K_S.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:89111d80ac38d2042ca034053bb3d3672487f9df88c17876775883d8fc6147b2
size 7255534112

3
gemma-3-12b-it-pruned-Q6_K.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a5c8b97dd15e68fe869d0a2ca31615b5fb626d7747effe690cb253c7955fcee0
size 9005038112

3
gemma-3-12b-it-pruned-Q8_0.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:4b82a3e945cd534fc6d4163e313dc87206e0c0c19fe79b392193bca86c47c992
size 10981440032

3
imatrix/imatrix-gemma-3-12b-it-medium.dat Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:86a497f01d0fb4d23caf5d7c6d836b866d55d50310ce757ef308abb0d811f27e
size 7433124

3
imatrix/imatrix-gemma-3-12b-it-micro.dat Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ccda143543ee58e40683497b2998b30b7843a2533574d9d01bedeecb1e742491
size 7433123

3
imatrix/imatrix-gemma-3-12b-it-small.dat Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f1dc7135f18f3a9607c994a3b408fa5532861d21c9f659953fe55f9f423a83b8
size 7433123

3
imatrix/imatrix-gemma-3-12b-it-tiny.dat Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:6870a2c2b916ef301a2b742067457d65e28511d8e322ad60329d215cff89e064
size 7433122

15
logits/README.md Normal file
View File

@@ -0,0 +1,15 @@
# Split logits file
Due to upload size limits imposed by Hugging Face, the logits file has been split into three parts, and you must recreate the original file before being able to use it.
There are several ways to do this, but the simplest is:
### Unix/Linux/MacOS
```sh
cat gemma-3-12b-it-F16-0001-0003.logits gemma-3-12b-it-F16-0002-0003.logits gemma-3-12b-it-F16-0003-0003.logits > gemma-3-12b-it-F16.logits
```
### Windows
```sh
copy gemma-3-12b-it-F16-0001-0003.logits + gemma-3-12b-it-F16-0002-0003.logits + gemma-3-12b-it-F16-0003-0003.logits gemma-3-12b-it-F16.logits /b
```

3
logits/gemma-3-12b-it-F16-0001-0003.logits Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:88b04c61b089e06074d34b95771f4b9ee602dcd181b8e7bbc4806999aad6a6be
size 26843545600

3
logits/gemma-3-12b-it-F16-0002-0003.logits Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0bcf5b9cc24ad26b941f440faf5efd16fc867af0da32b7e98c775c9e784d19e3
size 26843545600

3
logits/gemma-3-12b-it-F16-0003-0003.logits Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:448e9f91b32be38b4f10c12424e08f67b28ef980b0c90ab89cc6e972d9d33a61
size 23423349780

13
scores/gemma-3-12b-it-F16.arc Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 40 key-value pairs and 626 tensors from ./gemma-3-12b-it-F16.gguf (version GGUF V3 (latest))
Final result: 69.2000 +/- 1.6869
Random chance: 25.0083 +/- 1.5824
llama_perf_context_print: load time = 11768.83 ms
llama_perf_context_print: prompt eval time = 74484.60 ms / 35749 tokens ( 2.08 ms per token, 479.95 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 75938.16 ms / 35750 tokens
ggml_metal_free: deallocating

12
scores/gemma-3-12b-it-F16.hsw Normal file
View File

@@ -0,0 +1,12 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 40 key-value pairs and 626 tensors from ./gemma-3-12b-it-F16.gguf (version GGUF V3 (latest))
750 81.20000000% [78.2474%, 83.8346%]
llama_perf_context_print: load time = 1473.70 ms
llama_perf_context_print: prompt eval time = 281267.73 ms / 125814 tokens ( 2.24 ms per token, 447.31 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 287221.72 ms / 125815 tokens
ggml_metal_free: deallocating

1813
scores/gemma-3-12b-it-F16.md Normal file
View File

File diff suppressed because it is too large Load Diff

13
scores/gemma-3-12b-it-F16.mmlu Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 40 key-value pairs and 626 tensors from ./gemma-3-12b-it-F16.gguf (version GGUF V3 (latest))
Final result: 45.3333 +/- 1.8190
Random chance: 25.0000 +/- 1.5822
llama_perf_context_print: load time = 1708.63 ms
llama_perf_context_print: prompt eval time = 168929.92 ms / 67555 tokens ( 2.50 ms per token, 399.90 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 171048.28 ms / 67556 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-F16.tqa Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 40 key-value pairs and 626 tensors from ./gemma-3-12b-it-F16.gguf (version GGUF V3 (latest))
Final result: 41.4667 +/- 1.8002
Random chance: 19.8992 +/- 1.4588
llama_perf_context_print: load time = 1540.31 ms
llama_perf_context_print: prompt eval time = 245844.98 ms / 50067 tokens ( 4.91 ms per token, 203.65 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 248430.74 ms / 50068 tokens
ggml_metal_free: deallocating

11
scores/gemma-3-12b-it-F16.wng Normal file
View File

@@ -0,0 +1,11 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 40 key-value pairs and 626 tensors from ./gemma-3-12b-it-F16.gguf (version GGUF V3 (latest))
Final Winogrande score(750 tasks): 74.8000 +/- 1.5864
llama_perf_context_print: load time = 1465.03 ms
llama_perf_context_print: prompt eval time = 58933.69 ms / 21857 tokens ( 2.70 ms per token, 370.87 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 59720.41 ms / 21858 tokens
ggml_metal_free: deallocating

1746
scores/gemma-3-12b-it-IQ3_M.md Normal file
View File

File diff suppressed because it is too large Load Diff

1746
scores/gemma-3-12b-it-IQ3_S.md Normal file
View File

File diff suppressed because it is too large Load Diff

1746
scores/gemma-3-12b-it-IQ4_NL.md Normal file
View File

File diff suppressed because it is too large Load Diff

1746
scores/gemma-3-12b-it-Q3_K_L.md Normal file
View File

File diff suppressed because it is too large Load Diff

1746
scores/gemma-3-12b-it-Q3_K_M.md Normal file
View File

File diff suppressed because it is too large Load Diff

1746
scores/gemma-3-12b-it-Q3_K_S.md Normal file
View File

File diff suppressed because it is too large Load Diff

1746
scores/gemma-3-12b-it-Q4_K_M.md Normal file
View File

File diff suppressed because it is too large Load Diff

1746
scores/gemma-3-12b-it-Q4_K_S.md Normal file
View File

File diff suppressed because it is too large Load Diff

1746
scores/gemma-3-12b-it-Q5_K_M.md Normal file
View File

File diff suppressed because it is too large Load Diff

1746
scores/gemma-3-12b-it-Q5_K_S.md Normal file
View File

File diff suppressed because it is too large Load Diff

1746
scores/gemma-3-12b-it-Q6_K.md Normal file
View File

File diff suppressed because it is too large Load Diff

1746
scores/gemma-3-12b-it-Q8_0.md Normal file
View File

File diff suppressed because it is too large Load Diff

13
scores/gemma-3-12b-it-iq3_m.arc Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ3_M.gguf (version GGUF V3 (latest))
Final result: 66.9333 +/- 1.7190
Random chance: 25.0083 +/- 1.5824
llama_perf_context_print: load time = 2521.51 ms
llama_perf_context_print: prompt eval time = 87825.27 ms / 35749 tokens ( 2.46 ms per token, 407.05 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 89287.36 ms / 35750 tokens
ggml_metal_free: deallocating

12
scores/gemma-3-12b-it-iq3_m.hsw Normal file
View File

@@ -0,0 +1,12 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ3_M.gguf (version GGUF V3 (latest))
750 74.66666667% [71.4337%, 77.6482%]
llama_perf_context_print: load time = 520.21 ms
llama_perf_context_print: prompt eval time = 302150.78 ms / 125814 tokens ( 2.40 ms per token, 416.39 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 308129.94 ms / 125815 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-iq3_m.mmlu Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ3_M.gguf (version GGUF V3 (latest))
Final result: 42.0000 +/- 1.8034
Random chance: 25.0000 +/- 1.5822
llama_perf_context_print: load time = 518.01 ms
llama_perf_context_print: prompt eval time = 157589.37 ms / 67555 tokens ( 2.33 ms per token, 428.68 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 159704.02 ms / 67556 tokens
ggml_metal_free: deallocating

37
scores/gemma-3-12b-it-iq3_m.ppx Normal file
View File

@@ -0,0 +1,37 @@
====== Perplexity statistics ======
Mean PPL(Q) : 35.478449 ± 0.411747
Mean PPL(base) : 9.042786 ± 0.071503
Cor(ln(PPL(Q)), ln(PPL(base))): 80.18%
Mean ln(PPL(Q)/PPL(base)) : 1.366958 ± 0.007075
Mean PPL(Q)/PPL(base) : 3.923398 ± 0.027759
Mean PPL(Q)-PPL(base) : 26.435664 ± 0.356985
====== KL divergence statistics ======
Mean KLD: 1.229937 ± 0.004277
Maximum KLD: 25.876987
99.9% KLD: 13.510958
99.0% KLD: 8.435304
99.0% KLD: 8.435304
Median KLD: 0.728065
10.0% KLD: 0.020280
5.0% KLD: 0.004677
1.0% KLD: 0.000320
Minimum KLD: -0.000003
====== Token probability statistics ======
Mean Δp: -6.562 ± 0.077 %
Maximum Δp: 99.860%
99.9% Δp: 90.203%
99.0% Δp: 66.591%
95.0% Δp: 36.644%
90.0% Δp: 19.461%
75.0% Δp: 1.507%
Median Δp: -0.332%
25.0% Δp: -9.977%
10.0% Δp: -45.830%
5.0% Δp: -75.458%
1.0% Δp: -98.595%
0.1% Δp: -99.993%
Minimum Δp: -100.000%
RMS Δp : 30.166 ± 0.089 %
Same top p: 65.259 ± 0.124 %

13
scores/gemma-3-12b-it-iq3_m.tqa Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ3_M.gguf (version GGUF V3 (latest))
Final result: 39.4667 +/- 1.7860
Random chance: 19.8992 +/- 1.4588
llama_perf_context_print: load time = 525.08 ms
llama_perf_context_print: prompt eval time = 127081.64 ms / 50067 tokens ( 2.54 ms per token, 393.98 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 129691.56 ms / 50068 tokens
ggml_metal_free: deallocating

11
scores/gemma-3-12b-it-iq3_m.wng Normal file
View File

@@ -0,0 +1,11 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ3_M.gguf (version GGUF V3 (latest))
Final Winogrande score(750 tasks): 67.4667 +/- 1.7119
llama_perf_context_print: load time = 522.07 ms
llama_perf_context_print: prompt eval time = 54601.23 ms / 21857 tokens ( 2.50 ms per token, 400.30 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 55373.40 ms / 21858 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-iq3_s.arc Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ3_S.gguf (version GGUF V3 (latest))
Final result: 65.0667 +/- 1.7420
Random chance: 25.0083 +/- 1.5824
llama_perf_context_print: load time = 2418.65 ms
llama_perf_context_print: prompt eval time = 88136.62 ms / 35749 tokens ( 2.47 ms per token, 405.61 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 89561.06 ms / 35750 tokens
ggml_metal_free: deallocating

12
scores/gemma-3-12b-it-iq3_s.hsw Normal file
View File

@@ -0,0 +1,12 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ3_S.gguf (version GGUF V3 (latest))
750 74.80000000% [71.5718%, 77.7755%]
llama_perf_context_print: load time = 510.06 ms
llama_perf_context_print: prompt eval time = 312098.17 ms / 125814 tokens ( 2.48 ms per token, 403.12 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 317986.17 ms / 125815 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-iq3_s.mmlu Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ3_S.gguf (version GGUF V3 (latest))
Final result: 42.8000 +/- 1.8079
Random chance: 25.0000 +/- 1.5822
llama_perf_context_print: load time = 499.63 ms
llama_perf_context_print: prompt eval time = 160758.88 ms / 67555 tokens ( 2.38 ms per token, 420.23 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 162859.07 ms / 67556 tokens
ggml_metal_free: deallocating

37
scores/gemma-3-12b-it-iq3_s.ppx Normal file
View File

@@ -0,0 +1,37 @@
====== Perplexity statistics ======
Mean PPL(Q) : 32.557804 ± 0.369351
Mean PPL(base) : 9.042786 ± 0.071503
Cor(ln(PPL(Q)), ln(PPL(base))): 80.53%
Mean ln(PPL(Q)/PPL(base)) : 1.281050 ± 0.006837
Mean PPL(Q)/PPL(base) : 3.600418 ± 0.024617
Mean PPL(Q)-PPL(base) : 23.515018 ± 0.314640
====== KL divergence statistics ======
Mean KLD: 1.196823 ± 0.004146
Maximum KLD: 24.545359
99.9% KLD: 12.464450
99.0% KLD: 8.054839
99.0% KLD: 8.054839
Median KLD: 0.702317
10.0% KLD: 0.019899
5.0% KLD: 0.004923
1.0% KLD: 0.000363
Minimum KLD: -0.000003
====== Token probability statistics ======
Mean Δp: -6.314 ± 0.076 %
Maximum Δp: 99.585%
99.9% Δp: 90.128%
99.0% Δp: 66.718%
95.0% Δp: 36.821%
90.0% Δp: 19.531%
75.0% Δp: 1.614%
Median Δp: -0.338%
25.0% Δp: -9.890%
10.0% Δp: -44.374%
5.0% Δp: -73.795%
1.0% Δp: -98.369%
0.1% Δp: -99.991%
Minimum Δp: -100.000%
RMS Δp : 29.810 ± 0.088 %
Same top p: 65.550 ± 0.124 %

13
scores/gemma-3-12b-it-iq3_s.tqa Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ3_S.gguf (version GGUF V3 (latest))
Final result: 40.4000 +/- 1.7930
Random chance: 19.8992 +/- 1.4588
llama_perf_context_print: load time = 509.01 ms
llama_perf_context_print: prompt eval time = 125818.94 ms / 50067 tokens ( 2.51 ms per token, 397.93 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 128433.23 ms / 50068 tokens
ggml_metal_free: deallocating

11
scores/gemma-3-12b-it-iq3_s.wng Normal file
View File

@@ -0,0 +1,11 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ3_S.gguf (version GGUF V3 (latest))
Final Winogrande score(750 tasks): 66.1333 +/- 1.7292
llama_perf_context_print: load time = 501.85 ms
llama_perf_context_print: prompt eval time = 53507.05 ms / 21857 tokens ( 2.45 ms per token, 408.49 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 54292.47 ms / 21858 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-iq4_nl.arc Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ4_NL.gguf (version GGUF V3 (latest))
Final result: 68.2667 +/- 1.7007
Random chance: 25.0083 +/- 1.5824
llama_perf_context_print: load time = 2844.87 ms
llama_perf_context_print: prompt eval time = 91988.56 ms / 35749 tokens ( 2.57 ms per token, 388.62 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 93449.81 ms / 35750 tokens
ggml_metal_free: deallocating

12
scores/gemma-3-12b-it-iq4_nl.hsw Normal file
View File

@@ -0,0 +1,12 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ4_NL.gguf (version GGUF V3 (latest))
750 75.46666667% [72.2626%, 78.4112%]
llama_perf_context_print: load time = 559.13 ms
llama_perf_context_print: prompt eval time = 328969.37 ms / 125814 tokens ( 2.61 ms per token, 382.45 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 334877.24 ms / 125815 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-iq4_nl.mmlu Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ4_NL.gguf (version GGUF V3 (latest))
Final result: 42.6667 +/- 1.8072
Random chance: 25.0000 +/- 1.5822
llama_perf_context_print: load time = 571.85 ms
llama_perf_context_print: prompt eval time = 171936.29 ms / 67555 tokens ( 2.55 ms per token, 392.91 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 174058.79 ms / 67556 tokens
ggml_metal_free: deallocating

37
scores/gemma-3-12b-it-iq4_nl.ppx Normal file
View File

@@ -0,0 +1,37 @@
====== Perplexity statistics ======
Mean PPL(Q) : 36.879518 ± 0.436763
Mean PPL(base) : 9.042786 ± 0.071503
Cor(ln(PPL(Q)), ln(PPL(base))): 80.79%
Mean ln(PPL(Q)/PPL(base)) : 1.405689 ± 0.007174
Mean PPL(Q)/PPL(base) : 4.078336 ± 0.029260
Mean PPL(Q)-PPL(base) : 27.836733 ± 0.381333
====== KL divergence statistics ======
Mean KLD: 1.190922 ± 0.004215
Maximum KLD: 25.421314
99.9% KLD: 13.739526
99.0% KLD: 8.520005
99.0% KLD: 8.520005
Median KLD: 0.710480
10.0% KLD: 0.017773
5.0% KLD: 0.003969
1.0% KLD: 0.000258
Minimum KLD: -0.000003
====== Token probability statistics ======
Mean Δp: -6.063 ± 0.075 %
Maximum Δp: 99.690%
99.9% Δp: 88.494%
99.0% Δp: 65.535%
95.0% Δp: 36.662%
90.0% Δp: 19.570%
75.0% Δp: 1.644%
Median Δp: -0.273%
25.0% Δp: -9.153%
10.0% Δp: -43.591%
5.0% Δp: -73.479%
1.0% Δp: -98.555%
0.1% Δp: -99.995%
Minimum Δp: -100.000%
RMS Δp : 29.578 ± 0.088 %
Same top p: 66.655 ± 0.123 %

13
scores/gemma-3-12b-it-iq4_nl.tqa Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ4_NL.gguf (version GGUF V3 (latest))
Final result: 40.1333 +/- 1.7910
Random chance: 19.8992 +/- 1.4588
llama_perf_context_print: load time = 557.58 ms
llama_perf_context_print: prompt eval time = 133950.32 ms / 50067 tokens ( 2.68 ms per token, 373.77 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 136572.17 ms / 50068 tokens
ggml_metal_free: deallocating

11
scores/gemma-3-12b-it-iq4_nl.wng Normal file
View File

@@ -0,0 +1,11 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-IQ4_NL.gguf (version GGUF V3 (latest))
Final Winogrande score(750 tasks): 68.5333 +/- 1.6968
llama_perf_context_print: load time = 558.43 ms
llama_perf_context_print: prompt eval time = 56861.97 ms / 21857 tokens ( 2.60 ms per token, 384.39 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 57651.50 ms / 21858 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-q3_k_l.arc Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_L.gguf (version GGUF V3 (latest))
Final result: 67.8667 +/- 1.7063
Random chance: 25.0083 +/- 1.5824
llama_perf_context_print: load time = 2488.69 ms
llama_perf_context_print: prompt eval time = 93489.45 ms / 35749 tokens ( 2.62 ms per token, 382.39 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 94938.12 ms / 35750 tokens
ggml_metal_free: deallocating

12
scores/gemma-3-12b-it-q3_k_l.hsw Normal file
View File

@@ -0,0 +1,12 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_L.gguf (version GGUF V3 (latest))
750 75.73333333% [72.5391%, 78.6653%]
llama_perf_context_print: load time = 525.75 ms
llama_perf_context_print: prompt eval time = 338885.65 ms / 125814 tokens ( 2.69 ms per token, 371.26 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 344811.83 ms / 125815 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-q3_k_l.mmlu Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_L.gguf (version GGUF V3 (latest))
Final result: 42.5333 +/- 1.8065
Random chance: 25.0000 +/- 1.5822
llama_perf_context_print: load time = 523.91 ms
llama_perf_context_print: prompt eval time = 174640.97 ms / 67555 tokens ( 2.59 ms per token, 386.82 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 176748.74 ms / 67556 tokens
ggml_metal_free: deallocating

37
scores/gemma-3-12b-it-q3_k_l.ppx Normal file
View File

@@ -0,0 +1,37 @@
====== Perplexity statistics ======
Mean PPL(Q) : 33.537202 ± 0.371564
Mean PPL(base) : 9.042786 ± 0.071503
Cor(ln(PPL(Q)), ln(PPL(base))): 79.10%
Mean ln(PPL(Q)/PPL(base)) : 1.310688 ± 0.006832
Mean PPL(Q)/PPL(base) : 3.708725 ± 0.025338
Mean PPL(Q)-PPL(base) : 24.494416 ± 0.318026
====== KL divergence statistics ======
Mean KLD: 1.263077 ± 0.004253
Maximum KLD: 23.683029
99.9% KLD: 12.829600
99.0% KLD: 8.069357
99.0% KLD: 8.069357
Median KLD: 0.755642
10.0% KLD: 0.018898
5.0% KLD: 0.004577
1.0% KLD: 0.000375
Minimum KLD: -0.000003
====== Token probability statistics ======
Mean Δp: -8.200 ± 0.078 %
Maximum Δp: 99.800%
99.9% Δp: 90.394%
99.0% Δp: 66.393%
95.0% Δp: 34.806%
90.0% Δp: 16.907%
75.0% Δp: 0.806%
Median Δp: -0.550%
25.0% Δp: -12.684%
10.0% Δp: -50.926%
5.0% Δp: -77.962%
1.0% Δp: -98.305%
0.1% Δp: -99.990%
Minimum Δp: -100.000%
RMS Δp : 31.084 ± 0.088 %
Same top p: 62.253 ± 0.126 %

13
scores/gemma-3-12b-it-q3_k_l.tqa Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_L.gguf (version GGUF V3 (latest))
Final result: 40.2667 +/- 1.7920
Random chance: 19.8992 +/- 1.4588
llama_perf_context_print: load time = 525.48 ms
llama_perf_context_print: prompt eval time = 136845.27 ms / 50067 tokens ( 2.73 ms per token, 365.87 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 139475.22 ms / 50068 tokens
ggml_metal_free: deallocating

11
scores/gemma-3-12b-it-q3_k_l.wng Normal file
View File

@@ -0,0 +1,11 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_L.gguf (version GGUF V3 (latest))
Final Winogrande score(750 tasks): 67.0667 +/- 1.7172
llama_perf_context_print: load time = 528.70 ms
llama_perf_context_print: prompt eval time = 58406.45 ms / 21857 tokens ( 2.67 ms per token, 374.22 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 59199.95 ms / 21858 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-q3_k_m.arc Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_M.gguf (version GGUF V3 (latest))
Final result: 68.6667 +/- 1.6949
Random chance: 25.0083 +/- 1.5824
llama_perf_context_print: load time = 2459.83 ms
llama_perf_context_print: prompt eval time = 93086.65 ms / 35749 tokens ( 2.60 ms per token, 384.04 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 94546.19 ms / 35750 tokens
ggml_metal_free: deallocating

12
scores/gemma-3-12b-it-q3_k_m.hsw Normal file
View File

@@ -0,0 +1,12 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_M.gguf (version GGUF V3 (latest))
750 75.20000000% [71.9861%, 78.1570%]
llama_perf_context_print: load time = 506.32 ms
llama_perf_context_print: prompt eval time = 333957.19 ms / 125814 tokens ( 2.65 ms per token, 376.74 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 339911.25 ms / 125815 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-q3_k_m.mmlu Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_M.gguf (version GGUF V3 (latest))
Final result: 44.4000 +/- 1.8155
Random chance: 25.0000 +/- 1.5822
llama_perf_context_print: load time = 510.44 ms
llama_perf_context_print: prompt eval time = 172922.20 ms / 67555 tokens ( 2.56 ms per token, 390.67 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 175045.07 ms / 67556 tokens
ggml_metal_free: deallocating

37
scores/gemma-3-12b-it-q3_k_m.ppx Normal file
View File

@@ -0,0 +1,37 @@
====== Perplexity statistics ======
Mean PPL(Q) : 34.029948 ± 0.376082
Mean PPL(base) : 9.042786 ± 0.071503
Cor(ln(PPL(Q)), ln(PPL(base))): 78.62%
Mean ln(PPL(Q)/PPL(base)) : 1.325274 ± 0.006875
Mean PPL(Q)/PPL(base) : 3.763215 ± 0.025870
Mean PPL(Q)-PPL(base) : 24.987163 ± 0.322907
====== KL divergence statistics ======
Mean KLD: 1.300974 ± 0.004321
Maximum KLD: 23.795326
99.9% KLD: 12.854443
99.0% KLD: 8.215378
99.0% KLD: 8.215378
Median KLD: 0.782525
10.0% KLD: 0.019878
5.0% KLD: 0.004841
1.0% KLD: 0.000391
Minimum KLD: -0.000001
====== Token probability statistics ======
Mean Δp: -8.606 ± 0.079 %
Maximum Δp: 99.896%
99.9% Δp: 91.072%
99.0% Δp: 67.090%
95.0% Δp: 34.759%
90.0% Δp: 16.501%
75.0% Δp: 0.726%
Median Δp: -0.601%
25.0% Δp: -13.523%
10.0% Δp: -52.748%
5.0% Δp: -79.067%
1.0% Δp: -98.555%
0.1% Δp: -99.990%
Minimum Δp: -100.000%
RMS Δp : 31.537 ± 0.088 %
Same top p: 61.420 ± 0.127 %

13
scores/gemma-3-12b-it-q3_k_m.tqa Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_M.gguf (version GGUF V3 (latest))
Final result: 40.0000 +/- 1.7900
Random chance: 19.8992 +/- 1.4588
llama_perf_context_print: load time = 504.87 ms
llama_perf_context_print: prompt eval time = 134889.64 ms / 50067 tokens ( 2.69 ms per token, 371.17 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 137522.13 ms / 50068 tokens
ggml_metal_free: deallocating

11
scores/gemma-3-12b-it-q3_k_m.wng Normal file
View File

@@ -0,0 +1,11 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_M.gguf (version GGUF V3 (latest))
Final Winogrande score(750 tasks): 65.8667 +/- 1.7325
llama_perf_context_print: load time = 501.08 ms
llama_perf_context_print: prompt eval time = 56606.33 ms / 21857 tokens ( 2.59 ms per token, 386.12 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 57401.53 ms / 21858 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-q3_k_s.arc Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_S.gguf (version GGUF V3 (latest))
Final result: 67.6000 +/- 1.7100
Random chance: 25.0083 +/- 1.5824
llama_perf_context_print: load time = 2303.10 ms
llama_perf_context_print: prompt eval time = 91534.38 ms / 35749 tokens ( 2.56 ms per token, 390.55 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 92983.34 ms / 35750 tokens
ggml_metal_free: deallocating

12
scores/gemma-3-12b-it-q3_k_s.hsw Normal file
View File

@@ -0,0 +1,12 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_S.gguf (version GGUF V3 (latest))
750 75.20000000% [71.9861%, 78.1570%]
llama_perf_context_print: load time = 501.43 ms
llama_perf_context_print: prompt eval time = 325090.51 ms / 125814 tokens ( 2.58 ms per token, 387.01 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 331071.94 ms / 125815 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-q3_k_s.mmlu Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_S.gguf (version GGUF V3 (latest))
Final result: 43.7333 +/- 1.8126
Random chance: 25.0000 +/- 1.5822
llama_perf_context_print: load time = 508.75 ms
llama_perf_context_print: prompt eval time = 171219.64 ms / 67555 tokens ( 2.53 ms per token, 394.55 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 173347.99 ms / 67556 tokens
ggml_metal_free: deallocating

37
scores/gemma-3-12b-it-q3_k_s.ppx Normal file
View File

@@ -0,0 +1,37 @@
====== Perplexity statistics ======
Mean PPL(Q) : 36.103787 ± 0.402250
Mean PPL(base) : 9.042786 ± 0.071503
Cor(ln(PPL(Q)), ln(PPL(base))): 77.13%
Mean ln(PPL(Q)/PPL(base)) : 1.384430 ± 0.007124
Mean PPL(Q)/PPL(base) : 3.992551 ± 0.028445
Mean PPL(Q)-PPL(base) : 27.061001 ± 0.350072
====== KL divergence statistics ======
Mean KLD: 1.395578 ± 0.004559
Maximum KLD: 24.088387
99.9% KLD: 13.768757
99.0% KLD: 8.603606
99.0% KLD: 8.603606
Median KLD: 0.858334
10.0% KLD: 0.021285
5.0% KLD: 0.005174
1.0% KLD: 0.000393
Minimum KLD: -0.000002
====== Token probability statistics ======
Mean Δp: -9.212 ± 0.080 %
Maximum Δp: 99.844%
99.9% Δp: 91.248%
99.0% Δp: 67.876%
95.0% Δp: 34.787%
90.0% Δp: 16.139%
75.0% Δp: 0.593%
Median Δp: -0.696%
25.0% Δp: -14.997%
10.0% Δp: -54.930%
5.0% Δp: -79.831%
1.0% Δp: -98.568%
0.1% Δp: -99.992%
Minimum Δp: -100.000%
RMS Δp : 32.089 ± 0.088 %
Same top p: 59.830 ± 0.128 %

13
scores/gemma-3-12b-it-q3_k_s.tqa Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_S.gguf (version GGUF V3 (latest))
Final result: 39.8667 +/- 1.7890
Random chance: 19.8992 +/- 1.4588
llama_perf_context_print: load time = 501.19 ms
llama_perf_context_print: prompt eval time = 132421.97 ms / 50067 tokens ( 2.64 ms per token, 378.09 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 135070.83 ms / 50068 tokens
ggml_metal_free: deallocating

11
scores/gemma-3-12b-it-q3_k_s.wng Normal file
View File

@@ -0,0 +1,11 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q3_K_S.gguf (version GGUF V3 (latest))
Final Winogrande score(750 tasks): 66.5333 +/- 1.7242
llama_perf_context_print: load time = 501.82 ms
llama_perf_context_print: prompt eval time = 56390.15 ms / 21857 tokens ( 2.58 ms per token, 387.60 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 57175.02 ms / 21858 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-q4_k_m.arc Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q4_K_M.gguf (version GGUF V3 (latest))
Final result: 69.8667 +/- 1.6766
Random chance: 25.0083 +/- 1.5824
llama_perf_context_print: load time = 2958.46 ms
llama_perf_context_print: prompt eval time = 95508.61 ms / 35749 tokens ( 2.67 ms per token, 374.30 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 97132.91 ms / 35750 tokens
ggml_metal_free: deallocating

12
scores/gemma-3-12b-it-q4_k_m.hsw Normal file
View File

@@ -0,0 +1,12 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q4_K_M.gguf (version GGUF V3 (latest))
750 77.46666667% [74.3409%, 80.3125%]
llama_perf_context_print: load time = 567.97 ms
llama_perf_context_print: prompt eval time = 339481.51 ms / 125814 tokens ( 2.70 ms per token, 370.61 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 345440.37 ms / 125815 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-q4_k_m.mmlu Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q4_K_M.gguf (version GGUF V3 (latest))
Final result: 43.6000 +/- 1.8119
Random chance: 25.0000 +/- 1.5822
llama_perf_context_print: load time = 576.27 ms
llama_perf_context_print: prompt eval time = 176741.59 ms / 67555 tokens ( 2.62 ms per token, 382.22 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 178859.75 ms / 67556 tokens
ggml_metal_free: deallocating

37
scores/gemma-3-12b-it-q4_k_m.ppx Normal file
View File

@@ -0,0 +1,37 @@
====== Perplexity statistics ======
Mean PPL(Q) : 31.219358 ± 0.342580
Mean PPL(base) : 9.042786 ± 0.071503
Cor(ln(PPL(Q)), ln(PPL(base))): 80.83%
Mean ln(PPL(Q)/PPL(base)) : 1.239071 ± 0.006532
Mean PPL(Q)/PPL(base) : 3.452405 ± 0.022552
Mean PPL(Q)-PPL(base) : 22.176573 ± 0.287880
====== KL divergence statistics ======
Mean KLD: 1.106608 ± 0.003932
Maximum KLD: 20.797123
99.9% KLD: 12.310450
99.0% KLD: 7.690277
99.0% KLD: 7.690277
Median KLD: 0.642402
10.0% KLD: 0.017366
5.0% KLD: 0.004376
1.0% KLD: 0.000366
Minimum KLD: -0.000003
====== Token probability statistics ======
Mean Δp: -7.494 ± 0.075 %
Maximum Δp: 99.726%
99.9% Δp: 87.733%
99.0% Δp: 63.430%
95.0% Δp: 33.085%
90.0% Δp: 16.533%
75.0% Δp: 0.946%
Median Δp: -0.461%
25.0% Δp: -11.162%
10.0% Δp: -47.083%
5.0% Δp: -74.484%
1.0% Δp: -98.197%
0.1% Δp: -99.992%
Minimum Δp: -100.000%
RMS Δp : 29.739 ± 0.088 %
Same top p: 65.218 ± 0.124 %

13
scores/gemma-3-12b-it-q4_k_m.tqa Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q4_K_M.gguf (version GGUF V3 (latest))
Final result: 40.8000 +/- 1.7958
Random chance: 19.8992 +/- 1.4588
llama_perf_context_print: load time = 569.78 ms
llama_perf_context_print: prompt eval time = 138110.48 ms / 50067 tokens ( 2.76 ms per token, 362.51 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 140746.53 ms / 50068 tokens
ggml_metal_free: deallocating

11
scores/gemma-3-12b-it-q4_k_m.wng Normal file
View File

@@ -0,0 +1,11 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q4_K_M.gguf (version GGUF V3 (latest))
Final Winogrande score(750 tasks): 68.4000 +/- 1.6988
llama_perf_context_print: load time = 573.82 ms
llama_perf_context_print: prompt eval time = 59035.21 ms / 21857 tokens ( 2.70 ms per token, 370.24 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 59825.11 ms / 21858 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-q4_k_s.arc Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q4_K_S.gguf (version GGUF V3 (latest))
Final result: 69.8667 +/- 1.6766
Random chance: 25.0083 +/- 1.5824
llama_perf_context_print: load time = 2939.37 ms
llama_perf_context_print: prompt eval time = 94796.83 ms / 35749 tokens ( 2.65 ms per token, 377.11 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 96249.68 ms / 35750 tokens
ggml_metal_free: deallocating

12
scores/gemma-3-12b-it-q4_k_s.hsw Normal file
View File

@@ -0,0 +1,12 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q4_K_S.gguf (version GGUF V3 (latest))
750 77.33333333% [74.2021%, 80.1860%]
llama_perf_context_print: load time = 567.43 ms
llama_perf_context_print: prompt eval time = 336113.08 ms / 125814 tokens ( 2.67 ms per token, 374.32 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 342078.71 ms / 125815 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-q4_k_s.mmlu Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q4_K_S.gguf (version GGUF V3 (latest))
Final result: 43.4667 +/- 1.8113
Random chance: 25.0000 +/- 1.5822
llama_perf_context_print: load time = 557.41 ms
llama_perf_context_print: prompt eval time = 175467.57 ms / 67555 tokens ( 2.60 ms per token, 385.00 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 177580.76 ms / 67556 tokens
ggml_metal_free: deallocating

37
scores/gemma-3-12b-it-q4_k_s.ppx Normal file
View File

@@ -0,0 +1,37 @@
====== Perplexity statistics ======
Mean PPL(Q) : 31.245217 ± 0.343080
Mean PPL(base) : 9.042786 ± 0.071503
Cor(ln(PPL(Q)), ln(PPL(base))): 80.86%
Mean ln(PPL(Q)/PPL(base)) : 1.239899 ± 0.006533
Mean PPL(Q)/PPL(base) : 3.455265 ± 0.022573
Mean PPL(Q)-PPL(base) : 22.202432 ± 0.288349
====== KL divergence statistics ======
Mean KLD: 1.105266 ± 0.003926
Maximum KLD: 20.987593
99.9% KLD: 12.310256
99.0% KLD: 7.698290
99.0% KLD: 7.698290
Median KLD: 0.643550
10.0% KLD: 0.017519
5.0% KLD: 0.004389
1.0% KLD: 0.000363
Minimum KLD: -0.000002
====== Token probability statistics ======
Mean Δp: -7.460 ± 0.075 %
Maximum Δp: 99.266%
99.9% Δp: 87.347%
99.0% Δp: 63.610%
95.0% Δp: 33.077%
90.0% Δp: 16.572%
75.0% Δp: 0.938%
Median Δp: -0.465%
25.0% Δp: -11.152%
10.0% Δp: -46.828%
5.0% Δp: -74.300%
1.0% Δp: -98.187%
0.1% Δp: -99.991%
Minimum Δp: -100.000%
RMS Δp : 29.698 ± 0.088 %
Same top p: 65.245 ± 0.124 %

13
scores/gemma-3-12b-it-q4_k_s.tqa Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q4_K_S.gguf (version GGUF V3 (latest))
Final result: 40.8000 +/- 1.7958
Random chance: 19.8992 +/- 1.4588
llama_perf_context_print: load time = 580.69 ms
llama_perf_context_print: prompt eval time = 137720.71 ms / 50067 tokens ( 2.75 ms per token, 363.54 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 140323.32 ms / 50068 tokens
ggml_metal_free: deallocating

11
scores/gemma-3-12b-it-q4_k_s.wng Normal file
View File

@@ -0,0 +1,11 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q4_K_S.gguf (version GGUF V3 (latest))
Final Winogrande score(750 tasks): 69.6000 +/- 1.6807
llama_perf_context_print: load time = 571.52 ms
llama_perf_context_print: prompt eval time = 58871.14 ms / 21857 tokens ( 2.69 ms per token, 371.27 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 59658.60 ms / 21858 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-q5_k_m.arc Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q5_K_M.gguf (version GGUF V3 (latest))
Final result: 68.8000 +/- 1.6929
Random chance: 25.0083 +/- 1.5824
llama_perf_context_print: load time = 3504.83 ms
llama_perf_context_print: prompt eval time = 95720.53 ms / 35749 tokens ( 2.68 ms per token, 373.47 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 97175.56 ms / 35750 tokens
ggml_metal_free: deallocating

12
scores/gemma-3-12b-it-q5_k_m.hsw Normal file
View File

@@ -0,0 +1,12 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q5_K_M.gguf (version GGUF V3 (latest))
750 77.86666667% [74.7577%, 80.6916%]
llama_perf_context_print: load time = 627.41 ms
llama_perf_context_print: prompt eval time = 342154.54 ms / 125814 tokens ( 2.72 ms per token, 367.71 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 348106.77 ms / 125815 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-q5_k_m.mmlu Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q5_K_M.gguf (version GGUF V3 (latest))
Final result: 45.3333 +/- 1.8190
Random chance: 25.0000 +/- 1.5822
llama_perf_context_print: load time = 644.49 ms
llama_perf_context_print: prompt eval time = 177127.61 ms / 67555 tokens ( 2.62 ms per token, 381.39 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 179250.91 ms / 67556 tokens
ggml_metal_free: deallocating

37
scores/gemma-3-12b-it-q5_k_m.ppx Normal file
View File

@@ -0,0 +1,37 @@
====== Perplexity statistics ======
Mean PPL(Q) : 30.668770 ± 0.335889
Mean PPL(base) : 9.042786 ± 0.071503
Cor(ln(PPL(Q)), ln(PPL(base))): 81.60%
Mean ln(PPL(Q)/PPL(base)) : 1.221278 ± 0.006415
Mean PPL(Q)/PPL(base) : 3.391518 ± 0.021755
Mean PPL(Q)-PPL(base) : 21.625984 ± 0.280606
====== KL divergence statistics ======
Mean KLD: 1.039101 ± 0.003848
Maximum KLD: 22.598866
99.9% KLD: 12.543374
99.0% KLD: 7.697062
99.0% KLD: 7.697062
Median KLD: 0.596171
10.0% KLD: 0.017018
5.0% KLD: 0.004340
1.0% KLD: 0.000363
Minimum KLD: -0.000000
====== Token probability statistics ======
Mean Δp: -7.245 ± 0.074 %
Maximum Δp: 99.814%
99.9% Δp: 85.186%
99.0% Δp: 61.478%
95.0% Δp: 32.505%
90.0% Δp: 16.393%
75.0% Δp: 0.971%
Median Δp: -0.442%
25.0% Δp: -10.578%
10.0% Δp: -45.165%
5.0% Δp: -72.951%
1.0% Δp: -98.402%
0.1% Δp: -99.994%
Minimum Δp: -100.000%
RMS Δp : 29.177 ± 0.088 %
Same top p: 66.780 ± 0.123 %

13
scores/gemma-3-12b-it-q5_k_m.tqa Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q5_K_M.gguf (version GGUF V3 (latest))
Final result: 41.2000 +/- 1.7984
Random chance: 19.8992 +/- 1.4588
llama_perf_context_print: load time = 653.09 ms
llama_perf_context_print: prompt eval time = 137695.68 ms / 50067 tokens ( 2.75 ms per token, 363.61 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 140323.22 ms / 50068 tokens
ggml_metal_free: deallocating

11
scores/gemma-3-12b-it-q5_k_m.wng Normal file
View File

@@ -0,0 +1,11 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q5_K_M.gguf (version GGUF V3 (latest))
Final Winogrande score(750 tasks): 68.8000 +/- 1.6929
llama_perf_context_print: load time = 647.06 ms
llama_perf_context_print: prompt eval time = 58456.53 ms / 21857 tokens ( 2.67 ms per token, 373.90 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 59248.75 ms / 21858 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-q5_k_s.arc Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q5_K_S.gguf (version GGUF V3 (latest))
Final result: 68.9333 +/- 1.6909
Random chance: 25.0083 +/- 1.5824
llama_perf_context_print: load time = 3403.68 ms
llama_perf_context_print: prompt eval time = 94455.25 ms / 35749 tokens ( 2.64 ms per token, 378.48 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 95912.08 ms / 35750 tokens
ggml_metal_free: deallocating

12
scores/gemma-3-12b-it-q5_k_s.hsw Normal file
View File

@@ -0,0 +1,12 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q5_K_S.gguf (version GGUF V3 (latest))
750 77.73333333% [74.6188%, 80.5653%]
llama_perf_context_print: load time = 627.98 ms
llama_perf_context_print: prompt eval time = 334461.40 ms / 125814 tokens ( 2.66 ms per token, 376.17 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 340403.29 ms / 125815 tokens
ggml_metal_free: deallocating

13
scores/gemma-3-12b-it-q5_k_s.mmlu Normal file
View File

@@ -0,0 +1,13 @@
build: 5540 (291f2b69) with Apple clang version 17.0.0 (clang-1700.0.13.3) for arm64-apple-darwin24.4.0
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 49151 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 600 tensors from ./gemma-3-12b-it-Q5_K_S.gguf (version GGUF V3 (latest))
Final result: 45.0667 +/- 1.8180
Random chance: 25.0000 +/- 1.5822
llama_perf_context_print: load time = 666.76 ms
llama_perf_context_print: prompt eval time = 174509.11 ms / 67555 tokens ( 2.58 ms per token, 387.11 tokens per second)
llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second)
llama_perf_context_print: total time = 176638.86 ms / 67556 tokens
ggml_metal_free: deallocating

37
scores/gemma-3-12b-it-q5_k_s.ppx Normal file
View File

@@ -0,0 +1,37 @@
====== Perplexity statistics ======
Mean PPL(Q) : 30.644285 ± 0.335716
Mean PPL(base) : 9.042786 ± 0.071503
Cor(ln(PPL(Q)), ln(PPL(base))): 81.62%
Mean ln(PPL(Q)/PPL(base)) : 1.220479 ± 0.006413
Mean PPL(Q)/PPL(base) : 3.388810 ± 0.021733
Mean PPL(Q)-PPL(base) : 21.601499 ± 0.280413
====== KL divergence statistics ======
Mean KLD: 1.038541 ± 0.003836
Maximum KLD: 22.424044
99.9% KLD: 12.381549
99.0% KLD: 7.687484
99.0% KLD: 7.687484
Median KLD: 0.597444
10.0% KLD: 0.016928
5.0% KLD: 0.004340
1.0% KLD: 0.000359
Minimum KLD: -0.000003
====== Token probability statistics ======
Mean Δp: -7.203 ± 0.074 %
Maximum Δp: 99.718%
99.9% Δp: 84.842%
99.0% Δp: 61.523%
95.0% Δp: 32.607%
90.0% Δp: 16.381%
75.0% Δp: 0.982%
Median Δp: -0.437%
25.0% Δp: -10.550%
10.0% Δp: -45.107%
5.0% Δp: -72.871%
1.0% Δp: -98.353%
0.1% Δp: -99.993%
Minimum Δp: -100.000%
RMS Δp : 29.119 ± 0.088 %
Same top p: 66.765 ± 0.123 %

Some files were not shown because too many files have changed in this diff Show More