---
base_model: Qwen/Qwen3.5-4B
library_name: transformers
pipeline_tag: text-generation
language:
- en
tags:
- automaster
- skill-ft
- llm-agents
- tool-use
- arxiv:2611.26510
---
# AutoMaster-bc-4b
**AutoMaster-bc-4b is the 4B AutoMaster checkpoint for multi-step web search and evidence-based question answering.** Based on **Qwen3.5-4B**, it is trained with **Skill-Guided Fine-Tuning (Skill-FT)** on demonstrations produced by a skill-guided teacher.
## Model Overview
| Property | Description |
| --- | --- |
| Base model | [Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) |
| Domain | Search |
| Benchmark | BrowseComp (Lite-185) |
| Training | Full-parameter supervised fine-tuning with Skill-FT |
| Release format | Full model weights in Hugging Face Safetensors format |
| Inference framework | [AweAgent](https://github.com/RUCAIBox/AutoMaster/tree/main/recipes/inference) |
This repository contains the **fine-tuned student model**. Generated skills shape the teacher's training demonstrations and are absent during student evaluation.
## Method: Skill-Guided Fine-Tuning
Successful trajectories can still demonstrate weak habits, such as unsupported assumptions or superficial verification. AutoMaster uses an LLM agent to study trajectories, develop natural-language skills, and select demonstrations that teach more effective behavior.
- **Inner loop:** analyze behavior, design and test skills, and curate teacher demonstrations.
- **Outer loop:** fine-tune the student and use evaluation feedback to guide the next round.
*Framework figure from the [AutoMaster paper](https://arxiv.org/abs/2611.26510), distributed with the [public project repository](https://github.com/RUCAIBox/AutoMaster).*
## Training
The paper uses **512 training trajectories** and **three epochs of full-parameter SFT** per run. Skill instructions remain in the training context. The Vanilla SFT comparison uses the same teacher, task pool, training set size, and SFT recipe, with successful unguided teacher trajectories.
## Evaluation
Results on **BrowseComp (Lite-185)**, as reported in [Table 1 of the paper](https://arxiv.org/html/2611.26510v1#S4.T1):
| Model | Accuracy (%) ↑ (mean ± SD) |
| --- | ---: |
| Qwen3.5-4B | 33.69 ± 0.82 |
| Qwen3.5-4B + Vanilla SFT | 43.24 ± 0.54 |
| **AutoMaster-bc-4b** | **46.13 ± 2.98** |
Scores are the mean ± sample standard deviation over **three evaluation runs per checkpoint**. Evaluation uses the domain-specific AweAgent configuration, without generated skill prompts. See the [paper](https://arxiv.org/html/2611.26510v1) and [inference guide](https://github.com/RUCAIBox/AutoMaster/blob/main/recipes/inference/README.md) for task assets, tools, and evaluation settings.
## Using the Model
Use a Qwen3.5-compatible serving backend and keep the checkpoint's tokenizer and chat template with its weights. The public repository provides a serving script and benchmark configurations:
```bash
# Run from the AutoMaster repository after setting up its inference environment.
export MODEL_PATH="RUC-AIBOX/AutoMaster-bc-4b"
export SERVED_MODEL_NAME=automaster
bash recipes/inference/serve_qwen35.sh
```
In a separate terminal with AweAgent installed and the benchmark assets configured:
```bash
awe-agent run --config recipes/inference/browsecomp.yaml --max-instances 1
```
Follow the [inference guide](https://github.com/RUCAIBox/AutoMaster/blob/main/recipes/inference/README.md) to configure the endpoint and required tools or sandbox.
## Related Models
[bc-4b](https://huggingface.co/RUC-AIBOX/AutoMaster-bc-4b) · [bc-9b](https://huggingface.co/RUC-AIBOX/AutoMaster-bc-9b) · [doc2repo-4b](https://huggingface.co/RUC-AIBOX/AutoMaster-doc2repo-4b) · [doc2repo-9b](https://huggingface.co/RUC-AIBOX/AutoMaster-doc2repo-9b) · [tb21-4b](https://huggingface.co/RUC-AIBOX/AutoMaster-tb21-4b) · [tb21-9b](https://huggingface.co/RUC-AIBOX/AutoMaster-tb21-9b)
## Citation
```bibtex
@misc{chen2026automasterautonomousposttrainingmaking,
title={AutoMaster: Autonomous Post-Training by Making the Same Model a Better Teacher},
author={Guoxin Chen and Shuang Sun and Fanzhe Meng and Xinyu Tang and Ruihua Song and Wayne Xin Zhao},
year={2026},
eprint={2611.26510},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2611.26510}
}
```