69 lines
2.6 KiB
Markdown
69 lines
2.6 KiB
Markdown
---
|
||
license: apple-amlr
|
||
base_model:
|
||
- Qwen/Qwen3-4B-Thinking-2507
|
||
tags:
|
||
- self-distillation
|
||
- code-generation
|
||
library_name: transformers
|
||
---
|
||
|
||
# SimpleSD-4B-thinking
|
||
|
||
This model is an example of the **Simple Self-Distillation (SimpleSD)** method that improves code generation by fine-tuning a language model on its own sampled outputs—without rewards, verifiers, teacher models, or reinforcement learning. Please see the paper below for more information. This uses Qwen for initialization.
|
||
|
||
- **Self-distillation sampling:** temperature=1.1, top_p=0.95, top_k=20
|
||
- **Evaluation sampling:** temperature=0.7, top_p=0.95, top_k=20
|
||
|
||
paper: https://arxiv.org/abs/2604.01193
|
||
|
||
code: https://github.com/apple/ml-ssd
|
||
|
||
|
||
## Notes
|
||
- These are research checkpoints for reproducibility.
|
||
- They are not optimized Qwen releases.
|
||
- They don't represent a broader open-source model strategy.
|
||
|
||
## Usage
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
model = AutoModelForCausalLM.from_pretrained("apple/SimpleSD-4B-thinking")
|
||
tokenizer = AutoTokenizer.from_pretrained("apple/SimpleSD-4B-thinking")
|
||
```
|
||
|
||
## Method
|
||
|
||
SimpleSD samples solutions from the base model using non-unit temperature and top-k/top-p truncation, then fine-tunes on those samples via standard supervised learning. Despite its simplicity, SimpleSD yields large gains on competitive programming benchmarks, with improvements concentrating on harder problems. The mechanism traces to resolving a *precision–exploration conflict*: SimpleSD reshapes token distributions in a context-dependent way so that a single global decoding configuration becomes far more effective at evaluation time.
|
||
|
||
## Results
|
||
|
||
LiveCodeBench (%)
|
||
|
||
| Model | LCBv6 pass@1 | LCBv6 pass@5 | LCBv5 pass@1 | LCBv5 pass@5 |
|
||
|---|---|---|---|---|
|
||
| Qwen3-4B-Thinking-2507 (base) | 54.5 | 67.5 | 59.6 | 70.3 |
|
||
| **+ SimpleSD (this model)** | **57.8** (+3.3) | **71.4** (+3.9) | **63.1** (+3.5) | **74.7** (+4.4) |
|
||
|
||
## Paper
|
||
|
||
[**Embarrassingly Simple Self-Distillation Improves Code Generation**](https://arxiv.org/abs/2604.01193)
|
||
|
||
```bibtex
|
||
@misc{zhang2026embarrassinglysimpleselfdistillationimproves,
|
||
title={Embarrassingly Simple Self-Distillation Improves Code Generation},
|
||
author={Ruixiang Zhang and Richard He Bai and Huangjie Zheng and Navdeep Jaitly and Ronan Collobert and Yizhe Zhang},
|
||
year={2026},
|
||
eprint={2604.01193},
|
||
archivePrefix={arXiv},
|
||
primaryClass={cs.CL},
|
||
url={https://arxiv.org/abs/2604.01193},
|
||
}
|
||
```
|
||
|
||
|
||
## License
|
||
|
||
This model is released under the [Apple Machine Learning Research Model License](https://huggingface.co/apple/SimpleSD-4B-thinking/blob/main/LICENSE). |