---
license: mit
base_model:
  - WeiboAI/VibeThinker-3B
library_name: transformers
pipeline_tag: text-generation
tags:
  - unsloth
  - lora
  - trl
  - sft
---

# VibeThinker-3B-SFT-Fable5-Glint

A LoRA fine-tune of [`WeiboAI/VibeThinker-3B`](https://huggingface.co/WeiboAI/VibeThinker-3B), supervised fine-tuned on `ermiaazarkhalili/Fable-5-Glint-Clean` (private).

| | |
| --- | --- |
| **Base model** | [`WeiboAI/VibeThinker-3B`](https://huggingface.co/WeiboAI/VibeThinker-3B) |
| **Architecture** | `Qwen2ForCausalLM` |
| **Parameters** | 3.1B |
| **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) |

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "ermiaazarkhalili/VibeThinker-3B-SFT-Fable5-Glint"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype='auto', device_map='auto')

messages = [{"role": "user", "content": "Explain gradient checkpointing in two sentences."}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors='pt'
).to(model.device)

outputs = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## 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` |

## Held-out evaluation

Next-token accuracy on a deterministic held-out split of `ermiaazarkhalili/Fable-5-Glint-Clean` (n = 199 samples), scored against the base model [`WeiboAI/VibeThinker-3B`](https://huggingface.co/WeiboAI/VibeThinker-3B). Assistant tokens only; both models are scored identically.

| Metric | Base | This model | Δ |
| --- | ---: | ---: | ---: |
| Top-1 accuracy | 0.5309 | 0.6719 | +0.1411 |
| Top-5 accuracy | 0.7563 | 0.8947 | +0.1384 |

A delta measures how far fine-tuning moved this model from its own starting point; it is not a ranking against other models, which start from different baselines.

## Observed training loss

Measured from our SLURM logs for this configuration. These are training-loss
observations only — see the held-out evaluation above for measured accuracy.

| SLURM job | Steps | First loss | Final loss |
| --- | --- | --- | --- |
| `45987994` | 1,554 | 1.7458 | 1.0123 |
| `46021015` | 1,554 | 1.7457 | 1.0127 |

## 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`.*