---
license: apache-2.0
base_model: Qwen/Qwen3-0.6B
tags:
- code
- qwen3
- text-generation
- sft
- grpo
- trl
language:
- en
pipeline_tag: text-generation
library_name: transformers
datasets:
- microsoft/rStar-Coder
- open-r1/codeforces-cots
---

# Qwen3-0.6B-Sushi-Coder

<p align="center">
  <img src="Qwen3-0.6B-Sushi-Coder.png" alt="Qwen3-0.6B-Sushi-Coder" width="400">
</p>

A 0.6B code generation model fine-tuned from Qwen3-0.6B for Python code generation.

## Training

This model was trained in two stages:

1. **GRPO** using TRL with reward model based on test execution and formatting
2. **SFT** on [microsoft/rStar-Coder](https://huggingface.co/datasets/microsoft/rStar-Coder) and [open-r1/codeforces-cots](https://huggingface.co/datasets/open-r1/codeforces-cots)

Training was done on HuggingFace Jobs infrastructure using TRL.

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "bigatuna/Qwen3-0.6B-Sushi-Coder"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

messages = [
    {"role": "user", "content": "Write a Python function to check if a number is prime."}
]

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=512,
    temperature=0.6,
    top_p=0.95,
    do_sample=True
)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
```

## Recommended Parameters

| Setting | Value |
|---------|-------|
| Temperature | 0.6 |
| Top-p | 0.95 |
| Top-k | 20 |
| Max tokens | 2048 |

Note: Avoid greedy decoding (temperature=0) as it can cause repetition issues with Qwen3 models.

## Evaluation

| Model | pass@1 |
|-------|--------|
| Qwen/Qwen3-0.6B (base) | 20.1% |
| **Qwen3-0.6B-Sushi-Coder** | **29.3%** |

Note: The model scores slightly higher with prompt tuning and max_model_len=4000, but the results above represent baseline settings.

The evaluation can be reproduced with the following command:

```bash
lm_eval \
  --model vllm \
  --model_args pretrained=bigatuna/Qwen3-0.6B-Sushi-Coder,tensor_parallel_size=1,dtype=float16,gpu_memory_utilization=0.8,max_model_len=2048,trust_remote_code=True \
  --tasks humaneval \
  --batch_size 1 \
  --confirm_run_unsafe_code
```

Baseline comparison:

```bash
lm_eval \
  --model vllm \
  --model_args pretrained=Qwen/Qwen3-0.6B,tensor_parallel_size=1,dtype=float16,gpu_memory_utilization=0.8,max_model_len=2048,trust_remote_code=True \
  --tasks humaneval \
  --batch_size 1 \
  --confirm_run_unsafe_code
```

## Limitations

- Optimized for Python; other languages may have reduced quality
- Small model size limits complex reasoning
- May generate plausible but incorrect code for edge cases

## License

Apache 2.0