98 lines
3.2 KiB
Markdown
98 lines
3.2 KiB
Markdown
---
|
|
license: mit
|
|
base_model:
|
|
- WeiboAI/VibeThinker-3B
|
|
library_name: gguf
|
|
pipeline_tag: text-generation
|
|
tags:
|
|
- gguf
|
|
- llama.cpp
|
|
- quantized
|
|
- unsloth
|
|
- lora
|
|
- trl
|
|
- sft
|
|
---
|
|
|
|
# VibeThinker-3B-SFT-Fable5-Glint-GGUF
|
|
|
|
GGUF quantizations of a LoRA fine-tune of [`WeiboAI/VibeThinker-3B`](https://huggingface.co/WeiboAI/VibeThinker-3B), supervised fine-tuned on `ermiaazarkhalili/Fable-5-Glint-Clean` (private).
|
|
|
|
Quantized from [`ermiaazarkhalili/VibeThinker-3B-SFT-Fable5-Glint`](https://huggingface.co/ermiaazarkhalili/VibeThinker-3B-SFT-Fable5-Glint). See that repository for the full-precision weights.
|
|
|
|
| | |
|
|
| --- | --- |
|
|
| **Base model** | [`WeiboAI/VibeThinker-3B`](https://huggingface.co/WeiboAI/VibeThinker-3B) |
|
|
| **Training data** | `ermiaazarkhalili/Fable-5-Glint-Clean` (private) |
|
|
| **Method** | LoRA supervised fine-tuning via [Unsloth](https://github.com/unslothai/unsloth) + [TRL](https://github.com/huggingface/trl) |
|
|
| **License** | `mit` (inherited from the base model) |
|
|
|
|
## Available quantizations
|
|
|
|
| File | Size |
|
|
| --- | --- |
|
|
| `vibethinker-3b-sft-fable5-glint.q4_k_m.gguf` | 1.93 GB |
|
|
| `vibethinker-3b-sft-fable5-glint.q5_k_m.gguf` | 2.22 GB |
|
|
| `vibethinker-3b-sft-fable5-glint.q8_0.gguf` | 3.29 GB |
|
|
|
|
## Usage
|
|
|
|
### llama.cpp
|
|
|
|
```bash
|
|
huggingface-cli download ermiaazarkhalili/VibeThinker-3B-SFT-Fable5-Glint-GGUF vibethinker-3b-sft-fable5-glint.q4_k_m.gguf --local-dir .
|
|
llama-cli -m vibethinker-3b-sft-fable5-glint.q4_k_m.gguf -p "Explain gradient checkpointing in two sentences." -n 256
|
|
```
|
|
|
|
### Ollama
|
|
|
|
```bash
|
|
echo 'FROM ./vibethinker-3b-sft-fable5-glint.q4_k_m.gguf' > Modelfile
|
|
ollama create vibethinker-3b-sft-fable5-glint-gguf -f Modelfile
|
|
ollama run vibethinker-3b-sft-fable5-glint-gguf
|
|
```
|
|
|
|
## Training configuration
|
|
|
|
| Setting | Value |
|
|
| --- | --- |
|
|
| LoRA rank (r) | 16 |
|
|
| LoRA alpha | 16 |
|
|
| Learning rate | 0.0002 |
|
|
| Epochs | 3 |
|
|
| Effective batch size | 8 (2 x 4 grad accum) |
|
|
| Max sequence length | 4096 |
|
|
| Base precision | 4-bit (QLoRA) |
|
|
| Target modules | `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj` |
|
|
|
|
## Observed training loss
|
|
|
|
Measured from our SLURM logs for this configuration. These are training-loss
|
|
observations only — no downstream benchmark evaluation has been run on this
|
|
model, so they should not be read as a quality claim.
|
|
|
|
| SLURM job | Steps | First loss | Final loss |
|
|
| --- | --- | --- | --- |
|
|
| `45987994` | 1,554 | 3.3731 | 1.1936 |
|
|
| `46021015` | 1,554 | 3.3731 | 1.1904 |
|
|
|
|
## Limitations
|
|
|
|
- No benchmark evaluation has been run on this checkpoint. The only reported
|
|
numbers are training-loss observations.
|
|
- Inherits the biases, knowledge cutoff and failure modes of the base model.
|
|
- Fine-tuned on a single instruction-following dataset; behaviour outside that
|
|
distribution is untested.
|
|
- LoRA adapters were merged into the base weights, so the merged model cannot
|
|
be detached from this fine-tune.
|
|
|
|
## Reproducing
|
|
|
|
Trained by `notebooks/fable_distillation_vibethinker-3b_fable-glint_unsloth.ipynb`, executed non-interactively with
|
|
papermill on a SLURM H100 partition (Unsloth + TRL, LoRA).
|
|
|
|
---
|
|
|
|
*Card generated from the training run's own configuration and logs by*
|
|
*`scripts/generate_hub_model_card.py`.*
|