build-small-hackathon/figment-finetuned-model-archive
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity
Files
main
figment-finetuned-model-arc…/README.md
ModelHub XC 54ab68d505 初始化项目，由ModelHub XC社区提供模型 
Model: build-small-hackathon/figment-finetuned-model-archive
Source: Original Platform
2026-08-13 17:17:17 +08:00

355 lines
23 KiB
Markdown
Raw Permalink Blame History

---
base_model: nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16
base_model_relation: finetune
library_name: transformers
pipeline_tag: text-generation
license: other
license_name: nvidia-nemotron-open-model-license
license_link: >-
https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/
language:
- en
tags:
- figment
- lora
- peft
- gguf
- v1
- v2
- v3
- v4
- v5
- v6
- v7
- v8
- v9
- v10
- v11
- v12
- v13
- v14p
- protocol-navigation
- synthetic-data
- not-for-clinical-use
- base_model:finetune:nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16
---
# Figment Finetuned Model Archive
This repository archives early Figment local-model training artifacts for `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16`.
Figment is a prototype protocol-navigation aid for trained field responders working with synthetic or de-identified rural-clinic and disaster-response scenarios. It is designed to structure field notes, preserve deterministic red-flag rules, cite retrieved protocol cards, plan missing observations, draft responder checklists, and prepare SBAR-style handoffs.
The published artifacts include the `figment_sft_v1` pilot merged BF16 checkpoint from June 8, 2026, the `figment_sft_v2` merged BF16/GGUF checkpoint from June 9, 2026, the `figment_sft_v3` merged BF16/GGUF checkpoint from June 10, 2026, and the `figment_sft_v4` through `figment_sft_v14p` merged BF16/GGUF checkpoints from the June 11-13, 2026 field-workflow loop. The v1 pilot is retained for archival continuity, v2 improved raw configured-model behavior on the locked 50-case harness, v3 improved the field-holdout surface, v4 established the first archived field-workflow checkpoint, v5 is retained as a regression artifact, v6-v13 show the corrected field-workflow iteration path, and v14p plus its repair-union harness run is the strongest archived local field-workflow checkpoint in this repository.
## Contents
| Path | Contents | Notes |
| --- | --- | --- |
| `figment_sft_v1/pilot-20260608-merged-bf16/` | `figment_sft_v1` pilot adapter merged into the BF16 base with `peft.merge_and_unload(safe_merge=True)` | Full merged Hugging Face weights for the v1 pilot checkpoint. No v1 GGUF sidecar is archived in this repo. |
| `v2-20260609-merged-bf16/` | `figment_sft_v2` adapter merged into the BF16 base with `peft.merge_and_unload(safe_merge=True)` | Full merged Hugging Face weights for the v2 locked-harness checkpoint. |
| `v2-20260609-merged-bf16.gguf` | BF16 GGUF conversion of the v2 merged checkpoint | SHA-256: `281251bf326bfef219fe213cf01d7457164972ce2f99067b0ccc1fdb5821ea01`. |
| `v3-20260610-merged-bf16/` | `figment_sft_v3` adapter merged into the BF16 base with `peft.merge_and_unload(safe_merge=True)` | Full merged Hugging Face weights for the v3 field-workflow model. |
| `v3-20260610-merged-bf16.gguf` | BF16 GGUF conversion of the v3 merged checkpoint | SHA-256: `7ee6439f87d50af289136a345ee73e633e20035c79582f942f03f9331bb8a658`. |
| `v4-20260611-merged-bf16/` | `figment_sft_v4` adapter merged into the BF16 base with `peft.merge_and_unload(safe_merge=True)` | Full merged Hugging Face weights for the v4 field-workflow model. |
| `v4-20260611-merged-bf16.gguf` | BF16 GGUF conversion of the v4 merged checkpoint | SHA-256: `7e11f2295b101e9312f97075b8e48cabd8cc89539e92c8fa4218c4973aa31d8d`. |
| `figment_sft_v5/figment-sft-v5-lora-merged-bf16/` | `figment_sft_v5` adapter merged into the BF16 base with `peft.merge_and_unload(safe_merge=True)` | Full merged Hugging Face weights for the v5 field-workflow regression artifact. |
| `figment_sft_v5/figment-sft-v5-lora-merged-bf16.bf16.gguf` | BF16 GGUF conversion of the v5 merged checkpoint | Published LFS SHA-256: `c7f9b38d267c2ab2b791b613e0227ce3d057e61b57b568b16ca501f2e516379c`. |
| `figment_sft_v6/figment-sft-v6-lora-merged-bf16/` | `figment_sft_v6` adapter merged into the BF16 base with `peft.merge_and_unload(safe_merge=True)` | Full merged Hugging Face weights for the v6 field-workflow model. |
| `figment_sft_v6/figment-sft-v6-lora-merged-bf16.bf16.gguf` | BF16 GGUF conversion of the v6 merged checkpoint | Published LFS SHA-256: `92fb2bb4a8686230f050c1696e6df749fe49ec4d41221ab9100785afa7e34009`. |
| `figment_sft_v7/figment-sft-v7-lora-merged-bf16/` | `figment_sft_v7` adapter merged into the BF16 base with `peft.merge_and_unload(safe_merge=True)` | Full merged Hugging Face weights for the v7 field-workflow model. |
| `figment_sft_v7/figment-sft-v7-lora-merged-bf16.bf16.gguf` | BF16 GGUF conversion of the v7 merged checkpoint | Published LFS SHA-256: `d85f9dd7137453035ae8ec96bcee1998358ad5975bb9c842fe9b7a077c4002b9`. |
| `figment_sft_v8/figment-sft-v8-lora-merged-bf16/` | `figment_sft_v8` adapter merged into the BF16 base with `peft.merge_and_unload(safe_merge=True)` | Full merged Hugging Face weights for the v8 field-workflow model. |
| `figment_sft_v8/figment-sft-v8-lora-merged-bf16.bf16.gguf` | BF16 GGUF conversion of the v8 merged checkpoint | Published LFS SHA-256: `d45660834ce2f9229d0e43ed3ac6bd041dba876f54ff1cc384889b9594b5e78d`. |
| `figment_sft_v9/figment-sft-v9-lora-merged-bf16/` | `figment_sft_v9` adapter merged into the BF16 base with `peft.merge_and_unload(safe_merge=True)` | Full merged Hugging Face weights for the v9 field-workflow model. |
| `figment_sft_v9/figment-sft-v9-lora-merged-bf16.bf16.gguf` | BF16 GGUF conversion of the v9 merged checkpoint | Published LFS SHA-256: `79ec6bfb55895c90ed4188d9e4052730ac07f2f5c6fe49c5fd7ef44c7e0a7d16`. |
| `figment_sft_v10/figment-sft-v10-lora-merged-bf16/` | `figment_sft_v10` adapter merged into the BF16 base with `peft.merge_and_unload(safe_merge=True)` | Full merged Hugging Face weights for the v10 field-workflow model. |
| `figment_sft_v10/figment-sft-v10-lora-merged-bf16.bf16.gguf` | BF16 GGUF conversion of the v10 merged checkpoint | Published LFS SHA-256: `85bc2978be155e1cdf12b42c8ccf84e1c1b65ad2da6b463d7be726d33cbd31aa`. |
| `figment_sft_v11/figment-sft-v11-lora-merged-bf16/` | `figment_sft_v11` adapter merged into the BF16 base with `peft.merge_and_unload(safe_merge=True)` | Full merged Hugging Face weights for the v11 field-workflow model. |
| `figment_sft_v11/figment-sft-v11-lora-merged-bf16.bf16.gguf` | BF16 GGUF conversion of the v11 merged checkpoint | Published LFS SHA-256: `cb5c99e32660547941a681853c30eff47cd2a9aee837fbdd3ee17684b44d4fd2`. |
| `figment_sft_v12/figment-sft-v12-lora-merged-bf16/` | `figment_sft_v12` adapter merged into the BF16 base with `peft.merge_and_unload(safe_merge=True)` | Full merged Hugging Face weights for the v12 field-workflow model. |
| `figment_sft_v12/figment-sft-v12-lora-merged-bf16.bf16.gguf` | BF16 GGUF conversion of the v12 merged checkpoint | Published LFS SHA-256: `164ebf943919b4c27a54dbce3380bc156bbad3c6e893f1d185d35801eac015b7`. |
| `figment_sft_v13/figment-sft-v13-lora-merged-bf16/` | `figment_sft_v13` adapter merged into the BF16 base with `peft.merge_and_unload(safe_merge=True)` | Full merged Hugging Face weights for the v13 field-workflow model. |
| `figment_sft_v13/figment-sft-v13-lora-merged-bf16.bf16.gguf` | BF16 GGUF conversion of the v13 merged checkpoint | Published LFS SHA-256: `1cedcc48d2edf82f31ebd20d8885bdd7b72d07b8d551b19838394ba57a1f2e1e`. |
| `figment_sft_v14p/figment-sft-v14p-lora-merged-bf16/` | `figment_sft_v14p` adapter merged into the BF16 base with `peft.merge_and_unload(safe_merge=True)` | Full merged Hugging Face weights for the v14p field-workflow model. |
| `figment_sft_v14p/figment-sft-v14p-lora-merged-bf16.bf16.gguf` | BF16 GGUF conversion of the v14p merged checkpoint | Published LFS SHA-256: `53de48e5f7a7fa22af7a682686adcf6c0be7c5c1fe72f72ea39d80bd68333f72`. |
## Intended Use
Use this repo as an artifact archive for:
- reproducing the Modal train/merge/GGUF proof chain,
- comparing later Figment checkpoints against a known early baseline,
- inspecting the v1 pilot merged BF16 checkpoint,
- evaluating the v2 locked-harness protocol-navigation checkpoint,
- evaluating the v3 local/off-grid protocol-navigation checkpoint,
- evaluating the v4 local/off-grid field-workflow checkpoint,
- evaluating the v5 regression artifact and v6-v14p local/off-grid field-workflow checkpoints,
- debugging protocol-navigation behavior in synthetic or de-identified scenarios.
Do not use these artifacts for clinical care, autonomous triage, diagnosis, prescribing, medication dosing, or replacing local protocol or trained responder judgment.
## Model Details
- Base model: `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16`
- Base model revision observed during the project: `dfaf35de3e30f1867dd8dbc38a7fc9fb52d3914f`
- Model family: Nemotron 3 Nano 4B BF16, text generation
- Adapter method: PEFT LoRA
- LoRA rank: 16
- LoRA alpha: 32
- LoRA dropout: 0.05
- Target modules: `up_proj`, `in_proj`, `q_proj`, `k_proj`, `out_proj`, `v_proj`, `down_proj`, `o_proj`
- Max sequence length used for local 4B training: 16384
- Language: English
- Domain: synthetic field-clinic and disaster-response protocol navigation
## V1 Pilot Checkpoint
The v1 pilot artifact was trained as `figment_sft_v1` and merged from Modal checkpoint `/checkpoints/figment_sft_v1/pilot-20260608` into `/checkpoints/figment_sft_v1/pilot-20260608-merged-bf16`.
Archive summary:
- Artifact path: `figment_sft_v1/pilot-20260608-merged-bf16/`
- Merge method: `peft.merge_and_unload(safe_merge=True)`
- Merged dtype: BF16
- Base model: `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16`
- HF shard 1 LFS SHA-256: `aebcb7fd3126d0100cc7e78e58e0ed49ab29aad8f858f2c6149637aced9c699f`
- HF shard 2 LFS SHA-256: `8a1a7b48e647dd43cb7941a9e6a3f7a839326865034f626b2705645b0e29c830`
- Tokenizer LFS SHA-256: `623c34567aebb18582765289fbe23d901c62704d6518d71866e0e58db892b5b7`
- GGUF sidecar: not archived; no v1 GGUF cache was present in `figment-eval-results:/model_cache/figment_sft_v1`.
## V2 Checkpoint
The v2 artifact was trained as `figment_sft_v2` and merged from Modal checkpoint `/checkpoints/figment_sft_v2/figment-sft-v2-lora` into `/checkpoints/figment_sft_v2/figment-sft-v2-lora-merged-bf16`.
Training data and merge summary:
- Training rows: 1500
- Train rows: 1352
- Validation rows: 148
- Navigator-full rows: 1000
- Focused-repair rows: 500
- Train split SHA-256: `27233926a2bd9320418ff10b0c14f3885834adf2f48865ee469c939e2ffeb68a`
- Validation split SHA-256: `7964c75cd3940a8549e6b8b2ef15b4d5cd45e8607af8f77a4982ffe01116bfb4`
- Merge method: `peft.merge_and_unload(safe_merge=True)`
- Merged dtype: BF16
- Base model: `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16`
- Merged manifest SHA-256: `6885f758f30a76e798fac73ebedd64684f3287d6b459f2b625029b03031179dc`
- HF shard 1 SHA-256: `9e224445985294263fce0437f82e55d116e90f5f19a5b995d47ee5081ff97c63`
- HF shard 2 SHA-256: `758eb779adf5379fb96ea42c4c38cfc6de9dc3d53c4e3863a7aea15ccebae5ae`
- GGUF SHA-256: `281251bf326bfef219fe213cf01d7457164972ce2f99067b0ccc1fdb5821ea01`
The v2 local evaluation run was `local_4b_v2_lora_20260609T103344Z` on the locked 50-case local harness.
## V3 Checkpoint
The v3 artifact was trained as `figment_sft_v3` and merged from Modal checkpoint `/checkpoints/figment_sft_v3/figment-sft-v3-lora` into `/checkpoints/figment_sft_v3/figment-sft-v3-lora-merged-bf16`.
Training and merge summary:
- Training run: `700/700` optimizer steps
- Final eval loss: `0.04357146`
- Final train loss: `0.60960097`
- Merge method: `peft.merge_and_unload(safe_merge=True)`
- Merged dtype: BF16
- Base model: `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16`
- Merged manifest SHA-256: `d18e72fb258764321ec17abd687af7214a480f491f11d83cf64e38824dc4e510`
- GGUF SHA-256: `7ee6439f87d50af289136a345ee73e633e20035c79582f942f03f9331bb8a658`
The clean v3 field-holdout eval was the sequential run `local_4b_v3_lora_field_holdout_20260610T102450Z`, not the earlier parallel run that hit a llama.cpp KV/context-overflow failure mode.
## V4 Checkpoint
The v4 artifact was trained as `figment_sft_v4` and merged from Modal checkpoint `/checkpoints/figment_sft_v4/figment-sft-v4-lora` into `/checkpoints/figment_sft_v4/figment-sft-v4-lora-merged-bf16`.
Training data and merge summary:
- Training rows: 1650
- Train rows: 1482
- Validation rows: 168
- Navigator-full rows: 1500
- Focused-repair rows: 150
- Full corpus SHA-256: `ef7a7c9a6a99927ba72ce244e03a9da3ab86d3cf5dc70786703fb5f8bdf2a289`
- Train split SHA-256: `f869d79da9ef670bc6479f8321e51b1f48cb5a16423265f34893a08e7648676e`
- Validation split SHA-256: `3ff7668b8216d6fa0be770d6d9ed5f1a0b12965f9312d5210b510807538738d3`
- Merge method: `peft.merge_and_unload(safe_merge=True)`
- Merged dtype: BF16
- Base model: `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16`
- Merged manifest SHA-256: `6678c0ec3a28817dba22eb9e7c682b9961f04bbfc688d0f1bcd137afaf8c8c38`
- HF shard 1 SHA-256: `1d95889e945363adcd70a0be54bc29407d49e28bf7a2c0415e1732d81d64186c`
- HF shard 2 SHA-256: `2a2e27563e78981c130349feece291c976cf7d5384690c6327795eef6d08d4c0`
- GGUF SHA-256: `7e11f2295b101e9312f97075b8e48cabd8cc89539e92c8fa4218c4973aa31d8d`
The v4 full field-holdout evaluation run was `local_4b_finetuned_v4_field_holdout_20260611T011930Z`. A separate 50-case evidence run was `local_4b_finetuned_v4_evidence_20260611T0010Z`.
## V5 Checkpoint
The v5 artifact was trained as `figment_sft_v5` and merged from Modal checkpoint `/checkpoints/figment_sft_v5/figment-sft-v5-lora` into `/checkpoints/figment_sft_v5/figment-sft-v5-lora-merged-bf16`.
Training data and merge summary:
- Training rows: 1300
- Train rows: 1170
- Validation rows: 130
- Navigator-full rows: 1100
- Focused-repair rows: 200
- Full corpus SHA-256: `3abc2dcb1f972ee6f536c273de69f72abe9a42e402a3548c451e442a3fcd4535`
- Train split SHA-256: `08ad6b76e958249b50bece528e0b26f5d3ef090166d7e5e0d48ddc46101496c7`
- Validation split SHA-256: `54aadd55ab41f00880483ff0beb08c9602aae23933efcabd328d1769617fbc1a`
- Merge method: `peft.merge_and_unload(safe_merge=True)`
- Merged dtype: BF16
- Base model: `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16`
- GGUF LFS SHA-256: `c7f9b38d267c2ab2b791b613e0227ce3d057e61b57b568b16ca501f2e516379c`
The v5 field-holdout run was `figment_sft_v5_field_workflow_holdout_modal_gpu_20260611_h100_gguf`; it is retained as a regression artifact because it scored only `2/150` competence successes despite passing final JSON validation.
## V6 Checkpoint
The v6 artifact was trained as `figment_sft_v6` and merged from Modal checkpoint `/checkpoints/figment_sft_v6/figment-sft-v6-lora` into `/checkpoints/figment_sft_v6/figment-sft-v6-lora-merged-bf16`.
Training data and merge summary:
- Training rows: 2000
- Train rows: 1800
- Validation rows: 200
- Navigator-full rows: 1180
- Focused-repair rows: 820
- Full corpus SHA-256: `268cb36d0d36697006609f346b76c79dbf127f82837f5a1f76d47059b031c595`
- Train split SHA-256: `b750779104e80a8a92c86437f9515da7a4ab97bc866c1e87f4d95fca269ab9c2`
- Validation split SHA-256: `ca388117f77325a57c70af7d69145b429bd443a5ae134ce1ab419373154e25cf`
- Merge method: `peft.merge_and_unload(safe_merge=True)`
- Merged dtype: BF16
- Base model: `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16`
- GGUF LFS SHA-256: `92fb2bb4a8686230f050c1696e6df749fe49ec4d41221ab9100785afa7e34009`
The v6 field-holdout run was `figment_sft_v6_field_workflow_holdout_modal_gpu_20260611_h100_gguf`.
## V7 Checkpoint
The v7 artifact was trained as `figment_sft_v7` and merged from Modal checkpoint `/checkpoints/figment_sft_v7/figment-sft-v7-lora` into `/checkpoints/figment_sft_v7/figment-sft-v7-lora-merged-bf16`.
Training data and merge summary:
- Training rows: 2800
- Train rows: 2520
- Validation rows: 280
- Navigator-full rows: 1740
- Focused-repair rows: 1060
- Full corpus SHA-256: `b8bc3830beb38577047dbb2b9760aa2845234e25f41457fbfc5ce25bb6821ac0`
- Train split SHA-256: `283615b21446346a9090ad6d45e750f5812222625ddaa5d2a83a15f663cb7d04`
- Validation split SHA-256: `fe7b683f5007ff1f3eaac2632c9d407a8671d23c944b17b192eae964c0bbaa8d`
- Merge method: `peft.merge_and_unload(safe_merge=True)`
- Merged dtype: BF16
- Base model: `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16`
- GGUF LFS SHA-256: `d85f9dd7137453035ae8ec96bcee1998358ad5975bb9c842fe9b7a077c4002b9`
The v7 field-holdout run was `figment_sft_v7_field_workflow_holdout_modal_gpu_20260612_h100_gguf`.
## V8-V14p Checkpoints
The v8-v14p artifacts continue the corrected field-workflow training loop. Each checkpoint was merged from its Modal LoRA adapter into the same BF16 base with `peft.merge_and_unload(safe_merge=True)` and converted to BF16 GGUF for local llama.cpp evaluation.
| Version | Training rows | Train rows | Validation rows | Navigator rows | Focused-repair rows | Full corpus SHA-256 | Train split SHA-256 | Validation split SHA-256 | GGUF LFS SHA-256 |
| --- | ---: | ---: | ---: | ---: | ---: | --- | --- | --- | --- |
| V8 | 3200 | 2880 | 320 | 2140 | 1060 | `fbf2adb675d01c007f6defc0292d04574d671bd64cd112310771bb4f5161cecc` | `e4d81265d0d7d56443fd6afd91cd996c546e680b7e53e7200b09321c4bae56f5` | `4668f8f8aa558fe2e765feae91a82c661753da3836265e6485d1225629b55097` | `d45660834ce2f9229d0e43ed3ac6bd041dba876f54ff1cc384889b9594b5e78d` |
| V9 | 3600 | 3240 | 360 | 2540 | 1060 | `ceb106258d4149305582620b5c4c308a7aa5854b6125e2c1d14b0d98cf5bbd6b` | `b3556bae88e13f980b22509a5463e192556515cdaf868f65f16ef4db41079513` | `e2f4e13c516bea567e5f3501fc0b11143c100bca1afc65f907cb1ad27b211a85` | `79ec6bfb55895c90ed4188d9e4052730ac07f2f5c6fe49c5fd7ef44c7e0a7d16` |
| V10 | 4400 | 3960 | 440 | 3340 | 1060 | `6ba2a10a4f6afb3ba9a061ec966a68122b1c520b832b5e8e110de3900c2968bd` | `2497bca472e188d202939e9a729d338e8fe6f30913d96e2b396339d421f7de4d` | `256a1674930e57ccc7d511ea0825ef487a115bbf3c2aa79e7f2a4cc933c198fd` | `85bc2978be155e1cdf12b42c8ccf84e1c1b65ad2da6b463d7be726d33cbd31aa` |
| V11 | 5200 | 4680 | 520 | 4140 | 1060 | `867c5622aded6a73657e37f0a1468fb5edcfcc5c30c4d0e8eb7b5024a4786051` | `3e1606855dadfc0e67f4d45f4c98e729b697d095be2b351d0aa159c71f347eb3` | `970c8d00aeed2bec1bc069ae229ffc7865988d21f31dfc37de784cbd8b771b52` | `cb5c99e32660547941a681853c30eff47cd2a9aee837fbdd3ee17684b44d4fd2` |
| V12 | 4960 | 4464 | 496 | 3900 | 1060 | `9e7ba0caab6137be3bf9936b8a0cd2aa70679d467e3555d507ad5af063fb3a4e` | `fe009fcf471cc61ddeb7e7aa7d993ad5dd28d4c58238050751d42fcfd3b79098` | `bdbd3e51d0bf354da25b449de2d4164561f6f8d453cbbe5a196853b9eba40b23` | `164ebf943919b4c27a54dbce3380bc156bbad3c6e893f1d185d35801eac015b7` |
| V13 | 4465 | 4017 | 448 | 3405 | 1060 | `e7d5f55259c4a0cbfc81e16c31a8a374837c654ea8e5723434ac882ce835da2b` | `43106d3af0f494ca5ead39290f3ad142c7a1f73e46a98759a92aed7814083290` | `f16c98ea146a7a17785a761d50df00efbc5781b272085ca5885540c0a33a0645` | `1cedcc48d2edf82f31ebd20d8885bdd7b72d07b8d551b19838394ba57a1f2e1e` |
| V14p | 5335 | 4801 | 534 | 4275 | 1060 | `b455460870c70c2072491b754ed128e04cee7e63f4876cd6e6bacc92164788d9` | `378379eccba716001eb30a4bee05948a5bcb34ef2caa6801442be733c0f5fff6` | `aaf5d7c0b3236c98f7d29fcd9898ea6d6789978d468fb1e26d64b40097d2b86e` | `53de48e5f7a7fa22af7a682686adcf6c0be7c5c1fe72f72ea39d80bd68333f72` |
## Training Data
The model artifacts use synthetic and de-identified datasets generated inside the Figment project. Published training corpora are available in the dataset repository `build-small-hackathon/figment-eval-traces` under configs `figment_sft_v1` through `figment_sft_v14p`. The dataset files are not duplicated in this model repository.
The examples were synthetic. They were designed to teach Figment's harness behavior, not to store medical knowledge. They included full navigator outputs and focused repair tasks for schema, citations/pathways, SBAR handoff fields, missing observations, protocol urgency, and forbidden clinical language.
## Evaluation
For later eval-trace artifacts, see the dataset repository `build-small-hackathon/figment-eval-traces`.
Observed v2 locked-harness evaluation:
| Metric | V2 locked 50-case eval |
| --- | ---: |
| Total cases | 50 |
| Competence successes | 33/50 |
| Raw configured-model successes | 33/50 |
| Focused-repair successes | 0 |
| Full fallback uses | 0 |
| Final validation successes | 50/50 |
| Model-visible fields retained | 627/650 |
Observed v3 field-holdout evaluation:
| Metric | V3 field holdout |
| --- | ---: |
| Total cases | 150 |
| Competence successes | 107/150 |
| Raw configured-model successes | 93/150 |
| Focused-repair successes | 14 |
| Full fallback uses | 2 |
| Final validation successes | 148/150 |
| Model-visible fields retained | 1836/1950 |
Observed v4 evaluations:
| Metric | V4 50-case eval | V4 field holdout |
| --- | ---: | ---: |
| Total cases | 50 | 150 |
| Competence successes | 37/50 | 109/150 |
| Raw configured-model successes | 37/50 | 109/150 |
| Expected-label successes | 14/50 | 149/150 |
| Full fallback uses | 0 | 2 |
| Final validation successes | 50/50 | 148/150 |
| Model-visible fields retained | 624/650 | 1846/1950 |
Observed v5-v7 field-holdout evaluations:
| Metric | V5 field holdout | V6 field holdout | V7 field holdout |
| --- | ---: | ---: | ---: |
| Total cases | 150 | 150 | 150 |
| Competence successes | 2/150 | 142/150 | 148/150 |
| Raw configured-model successes | 2/150 | 142/150 | 148/150 |
| Expected-label successes | 150/150 | 146/150 | 145/150 |
| Full fallback uses | 0 | 0 | 0 |
| Final validation successes | 150/150 | 150/150 | 150/150 |
| Deterministic patch count | 302 | 21 | 4 |
| Model-visible field pass rate | 0.8451 | 0.9892 | 0.9979 |
| Mean latency | 4512.943 ms | 4407.568 ms | 4344.942 ms |
| P95 latency | 4714.603 ms | 4606.824 ms | 4565.243 ms |
Observed v8-v14p corrected field-holdout evaluations:
| Metric | V8 | V9 | V10 | V11 | V12 | V13 | V14p | V14p repair-union |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| Total cases | 150 | 150 | 150 | 150 | 150 | 150 | 150 | 150 |
| Competence successes | 146/150 | 146/150 | 147/150 | 145/150 | 146/150 | 146/150 | 146/150 | 150/150 |
| Raw configured-model successes | 146/150 | 146/150 | 147/150 | 143/150 | 146/150 | 145/150 | 146/150 | 146/150 |
| Expected-label successes | 150/150 | 150/150 | 150/150 | 148/150 | 150/150 | 149/150 | 150/150 | 150/150 |
| Full fallback uses | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| Final validation successes | 150/150 | 150/150 | 150/150 | 150/150 | 150/150 | 150/150 | 150/150 | 150/150 |
| Deterministic patch count | 8 | 8 | 6 | 23 | 8 | 15 | 8 | 0 |
| Model-visible field pass rate | 0.9959 | 0.9959 | 0.9969 | 0.9882 | 0.9959 | 0.9923 | 0.9959 | 1.0000 |
| Mean latency | 4282.523 ms | 4338.318 ms | 4341.864 ms | 4491.265 ms | 4932.291 ms | 4362.790 ms | 5142.904 ms | 4505.249 ms |
| P95 latency | 4477.092 ms | 4613.479 ms | 4602.941 ms | 4664.988 ms | 5181.336 ms | 4532.570 ms | 5642.153 ms | 4708.240 ms |
## Safety and Limitations
- Prototype only; not a medical device.
- Synthetic/de-identified scenarios only.
- The model must not diagnose, prescribe, dose medication, or autonomously triage.
- Deterministic red-flag rules and validators remain part of the Figment runtime. The model artifact alone is not the full safety system.
- Outputs require trained responder review and local protocol/supervisor/clinician judgment.
- The checkpoints may produce malformed, incomplete, unsupported, or overconfident outputs without the Figment harness.
## License and Attribution
This archive is derived from `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16` and is governed by the same upstream [NVIDIA Nemotron Open Model License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/). Review the upstream model card and license before reuse. The Figment application code is Apache-2.0, and Figment synthetic datasets are documented separately as CC-BY-4.0 where published.
## Citation
No paper is associated with these artifacts. Please cite the base model according to NVIDIA's guidance and cite this repository if using the Figment artifacts directly.
Reference in New Issue View Git Blame Copy Permalink