| --- |
| license: apple-amlr |
| base_model: |
| - Qwen/Qwen3-4B-Instruct-2507 |
| tags: |
| - self-distillation |
| - code-generation |
| library_name: transformers |
| --- |
| |
| # SimpleSD-4B-instruct |
|
|
| 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=1.6, top_p=0.8, top_k=20 |
| - **Evaluation sampling:** temperature=1.1, top_p=0.8, 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-instruct") |
| tokenizer = AutoTokenizer.from_pretrained("apple/SimpleSD-4B-instruct") |
| ``` |
|
|
| ## 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-Instruct-2507 (base) | 34.0 | 41.0 | 34.3 | 45.4 | |
| | **+ SimpleSD (this model)** | **41.5** (+7.5) | **56.8** (+15.8) | **45.7** (+11.4) | **61.9** (+16.5) | |
|
|
| ## 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-instruct/blob/main/LICENSE). |
