File size: 2,665 Bytes
1e64107 fc987ff 1e64107 d1f417f 1e64107 5309422 1e64107 fc987ff 1e64107 777470d b0a0ae1 d1f417f 1e64107 819ce71 1e64107 fc987ff 819ce71 fc987ff 1e64107 d1f417f 1e64107 d1f417f 1e64107 02e3aaf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 | ---
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). |