---
library_name: transformers
license: apache-2.0
base_model: Qwen/Qwen2.5-7B-Instruct
tags:
  - sdft
  - self-distillation
  - continual-learning
  - tool-use
  - qwen2.5
language:
  - en
pipeline_tag: text-generation
---

# Qwen2.5-7B-Instruct SDFT — Tool Use (Step 300)

This model is a **Self-Distillation Fine-Tuned (SDFT)** version of [Qwen/Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct), trained on the ToolAlpaca tool-use dataset.

SDFT is an on-policy learning method from ["Self-Distillation Enables Continual Learning"](https://arxiv.org/abs/2601.19897) that acquires new skills while preserving prior capabilities, significantly reducing catastrophic forgetting compared to standard SFT.

## Training Details

| Parameter | Value |
|-----------|-------|
| Base model | Qwen/Qwen2.5-7B-Instruct |
| Method | SDFT (On-Policy Self-Distillation) |
| Dataset | ToolAlpaca (4,046 training examples) |
| Training step | 300 / 1011 |
| Learning rate | 2e-5 (cosine schedule, 10% warmup) |
| Batch size | 32 (gradient accumulation) |
| Epochs | 1 |
| Precision | bf16 |
| Max prompt length | 1024 |
| Max completion length | 1024 |
| EMA alpha | 0.01 |
| Hardware | 1x NVIDIA L40S 48GB |
| Training time | ~42 hours (full run) |

## Evaluation Results

### Tool-Use Accuracy (ToolAlpaca test set, 68 examples)

| Metric | Base Model | This Model (Step 300) |
|--------|-----------|--------------------------|
| Greedy Accuracy | 54.4% | 44.1% |
| pass@1 | 52.6% | 44.3% |
| pass@5 | 61.5% | 59.0% |
| pass@10 | 64.3% | 64.3% |

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-300")
tokenizer = AutoTokenizer.from_pretrained("Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-300")

messages = [{"role": "user", "content": "Your tool-use prompt here"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## All Checkpoints

| Step | HuggingFace |
|------|-------------|
| 100 | [Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-100](https://huggingface.co/Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-100) |
| 200 | [Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-200](https://huggingface.co/Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-200) |
| 300 | [Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-300](https://huggingface.co/Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-300) |
| 400 | [Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-400](https://huggingface.co/Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-400) |
| 500 | [Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-500](https://huggingface.co/Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-500) |
| 600 | [Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-600](https://huggingface.co/Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-600) |
| 700 | [Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-700](https://huggingface.co/Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-700) |
| 800 | [Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-800](https://huggingface.co/Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-800) |
| 900 | [Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-900](https://huggingface.co/Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-900) |
| 1000 | [Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-1000](https://huggingface.co/Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-1000) |
| 1011 | [Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-1011](https://huggingface.co/Ayushnangia/qwen2.5-7b-instruct-sdft-tooluse-step-1011) |

## Citation

```bibtex
@article{shenfeld2025selfdistillation,
  title={Self-Distillation Enables Continual Learning},
  author={Shenfeld, Idan and others},
  journal={arXiv preprint arXiv:2601.19897},
  year={2025}
}
```