GAIA Capstone Agent
A small, inspectable browser agent. It runs Qwen 3 0.6B locally through Transformers.js/WebGPU, looks up supporting Wikipedia pages, reads supported benchmark attachments, and sends its answers to the official course evaluator only when you press Submit.
What runs where
Model weights and inference run in this browser; the project has no model API key and no paid Space compute. Audio attachments are transcribed locally with Whisper Tiny, and image attachments use LFM2.5-VL-450M when WebGPU is available. Questions and any evaluated answers are sent to the Hugging Face Agents Course scoring API when you submit. The leaderboard submission includes the HF username and this public Space code link, as required by the course.
The compact models may miss difficult questions or unsupported media. The benchmark uses exact matching. Review the generated answers before submitting.
Evaluation results
| # | Question | Agent answer | Tools |
|---|
Agent design
The agent combines a quantized Qwen model with deterministic handling for reverse-text and operation-table questions, Wikipedia search for research questions, in-browser spreadsheet parsing for workbook tasks, Whisper Tiny transcription for audio attachments, best-effort YouTube captions, and LFM2.5-VL image question answering when WebGPU is available. Caption access depends on a CORS-readable English track; unavailable captions are reported and the agent uses the question and other evidence. Files are fetched from the official course endpoint. Final output is normalized to a concise answer for the exact-match evaluator.
Models: Qwen 3 0.6B ONNX, Whisper Tiny English ONNX, and LFM2.5-VL-450M ONNX. Runtime: Transformers.js.
Source code: public Space repository · Course: Hugging Face Agents Course