README.md

---
license: apache-2.0
language:
  - en
  - zh
tags:
  - qwen3
  - qwen
  - causal-lm
  - transformers
  - instruction-tuning
  - sft
  - agents
  - code
library_name: transformers
pipeline_tag: text-generation
base_model: Qwen/Qwen3-8B
model_name: Qwen3-8B-ABC
---

<div align="center">
<h1 style="font-size: 40px; line-height: 1.1; margin: 0;">
  Qwen3-8B-ABC
</h1>
</div>

<br>

<div align="center">
💻 <a href="https://github.com/red-fox-yj/ABC-Bench"><b>Code</b></a>&nbsp&nbsp | &nbsp&nbsp
📑 <a href="https://arxiv.org/abs/2601.11077"><b>Paper</b></a>&nbsp&nbsp | &nbsp&nbsp
📝 <a href="https://dawning-road.github.io/blog/abc-bench"><b>Blog</b></a>&nbsp&nbsp | &nbsp&nbsp
🤗 <a href="https://huggingface.co/datasets/nex-agi/agent-sft"><b>Data</b></a>
</div>

**Qwen3-8B-ABC** is a supervised fine-tuned (SFT) variant of **Qwen/Qwen3-8B**, trained for **agentic backend coding** and **tool-using / instruction-following** behaviors.

## Model Details

- **Model name**: `Qwen3-8B-ABC`
- **Base model**: `Qwen/Qwen3-8B`
- **Model type**: Causal Language Model (decoder-only)
- **Training method**: Agentic Supervised Fine-Tuning (SFT)

## Training Data

This model was fine-tuned on [nex-agi/agent-sft](https://huggingface.co/datasets/nex-agi/agent-sft).

Please refer to the dataset card for detailed documentation, licensing, and usage constraints.

## Performance on ABC-Bench

Following the ABC-Bench paper’s evaluation protocol:

| Model | Setting | Average Pass@1 (%, 3 attempts) |
|---|---|:---:|
| Qwen3-8B-ABC | w/ SFT | **13.9%** |
| Qwen3-8B | w/o SFT | **8.3%** |

## Intended Use

Qwen3-8B-ABC is intended for:
- Agent-style instruction following for backend development tasks
- Code editing / patch generation in real repositories
- Command-line oriented debugging and step-by-step problem solving
- Research on automated software engineering and agent evaluation

## Usage

### Transformers (Python)

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "OpenMOSS-Team/Qwen3-8B-ABC"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

prompt = "Write a FastAPI endpoint that returns health status as JSON."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=True,
        temperature=0.7,
        top_p=0.9,
    )

print(tokenizer.decode(output[0], skip_special_tokens=True))
```

## Citation

```bibtex
@misc{yang2026abcbenchbenchmarkingagenticbackend,
      title={ABC-Bench: Benchmarking Agentic Backend Coding in Real-World Development}, 
      author={Jie Yang and Honglin Guo and Li Ji and Jiazheng Zhou and Rui Zheng and Zhikai Lei and Shuo Zhang and Zhiheng Xi and Shichun Liu and Yuxin Wang and Bo Wang and Yining Zheng and Tao Gui and Xipeng Qiu},
      year={2026},
      eprint={2601.11077},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2601.11077}, 
}
```

## Acknowledgements

- Base model: `Qwen/Qwen3-8B`
- Training dataset: `nex-agi/agent-sft`