---
base_model: Qwen/Qwen3.5-2B
language:
- en
pipeline_tag: image-text-to-text
library_name: transformers
tags:
- agents
- web
- sft
- qwen
---

<div align="center">

# A3-Qwen3.5-2B

| [**💾 Code**](https://github.com/McGill-NLP/agent-as-annotators) | [**📄 Paper**](https://huggingface.co/papers/2604.07776) | [**🌐 Website**](https://agent-as-annotators.github.io) |
| :--: | :--: | :--: |
| [**🤗 Dataset**](https://huggingface.co/datasets/McGill-NLP/A3-Synth) | [**🤖 Models**](https://huggingface.co/collections/McGill-NLP/a3-agent-as-annotators-69d854ab5b1993b10efc3fba) | [**📦 PyPI**](https://pypi.org/project/agent-as-annotators/) |

[**Structured Distillation of Web Agent Capabilities Enables Generalization**](https://huggingface.co/papers/2604.07776)

*Xing Han Lù, Siva Reddy*

</div>

**A3-Qwen3.5-2B** is a 2B multimodal web agent fine-tuned from [Qwen/Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) using the **Agent-as-Annotators (A3)** framework. It is trained on [A3-Synth](https://huggingface.co/datasets/McGill-NLP/A3-Synth), a dataset of high-quality synthetic trajectories generated through a structured teacher-student distillation process.

## Model Description
A3-Qwen3.5-2B is designed to navigate complex web environments by processing visual screenshots and text. By decomposing the synthetic data generation process into three modular roles—Task Designer, Annotator, and Supervisor—the A3 framework allows small, locally deployable models to achieve competitive performance on benchmarks like WebArena, even surpassing some larger closed-source models.

## Quick Start: Evaluation

You can evaluate the model using the `agent-as-annotators` toolkit:

### 1. Serve the model with vLLM
```bash
vllm serve --model McGill-NLP/A3-Qwen3.5-2B
```

### 2. Run evaluation
```bash
a3-eval --benchmark webarena_test --model A3-qwen3.5-2b
```

## Citation

If you find this model useful, please cite our work:

```bibtex
@misc{lu2025structured,
      title={Structured Distillation of Web Agent Capabilities Enables Generalization}, 
      author={Xing Han Lù and Siva Reddy},
      year={2025},
      eprint={2604.07776},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}
```