# qwen-2.5-instruct-sdft-science
This model is a fine-tuned version of [Qwen/Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct).
It has been trained using [TRL](https://github.com/huggingface/trl).

This model has been trained using [Self-Distillation Fine-Tuning](https://arxiv.org/abs/2601.19897) on the Released Science dataset.
Within this repo, you can find `/eval` directory, containing the evaluation results on the Tool Use evaluation split.
The model has been trained for 300 steps (the best checkopint we have obtained), scoring 64.5% on the evaluation set.

Our Reproduction Report of Tool Use: https://github.com/KickItLikeShika/sdft-reproduction-note

## Quick start

```python
from transformers import pipeline

question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
generator = pipeline("text-generation", model="KickItLikeShika/qwen-2.5-7b-instruct-sdft-science", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])
```

## Training procedure
W&B Report Tool Use Reproduction Report
https://api.wandb.ai/links/egyttsteam/d97ty5d9

### Framework versions

- TRL: 0.24.0
- Transformers: 4.57.1
- Pytorch: 2.9.0
- Datasets: 4.3.0
- Tokenizers: 0.22.2