Hugging Face Agents Course / Unit 4 project

Local reasoning.
Public evidence.
Human-reviewed submissions.

A Python research agent powered by a downloaded Qwen3 model. It gathers original sources, reads task attachments, and saves its tool traces before any score submission.

Qwen3 14B + Ollama No metered inference API Read-only research tools

This is the project documentation, not a hosted inference service. Run the agent on your own computer. This page does not run a model, send answers, or claim a passing score.

01 / INVESTIGATE

Find the evidence

Search public sources, read web pages and PDFs, and follow source links. Benchmark answer keys and solution repositories are excluded.

02 / REASON

Use bounded tools

Read spreadsheets and source text, calculate arithmetic, reverse text, and optionally transcribe audio locally. Python attachments are never executed.

03 / REVIEW

Keep control

Save answers and execution traces locally. Submission is a separate command requiring an explicit confirmation flag and a matching payload hash.

Reproduce locally

Install Python and Ollama, then download the project files from the public source tree. Run these commands from the project folder:

ollama pull qwen3:14b
python -m venv .venv
.venv\Scripts\python -m pip install -r requirements-lock.txt
.venv\Scripts\python -m unittest -v test_agent
.venv\Scripts\python run.py run --output runs --audio

The default quantized model download is approximately 9 GB. Adequate RAM and a supported GPU are recommended. Use requirements.txt and omit --audio if local speech transcription is not needed. The lock file includes the complete audio-enabled environment. For macOS or Linux, use the Python executable inside your virtual environment's bin directory.

Prepare a submission without sending it

.venv\Scripts\python run.py prepare --output runs --username YOUR_USERNAME --space YOUR_USERNAME/YOUR_SPACE

Review the generated submission.json and its SHA-256 hash. A separate submit command is available via python run.py submit --help. It writes to the official course leaderboard; do not run it without reviewing the destination, account, code link, and exact answers.

Limitations and assessment integrity

The agent does not include an answer lookup table. Search results and speech recognition can be inaccurate, sources may be inaccessible, and the text model cannot inspect images or watch videos. Unsupported questions may remain unanswered. Local model outputs are not an official score.

The course scoring service determines results. Obtain certificates only through the official certificate page after meeting its eligibility requirements.