用 Gemma 3 与 Ollama 统计英国学生黑客松赛事数据

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built

hackreport is a pipeline that reads public pages about past UK hackathons and reports on the student-led events season by season. For each event it records the organiser, whether it was student-led, and where available the size and number of projects.

I built it for a colleague of mine who was interested in seeing statistics of events across the year.
The problem: there is no single place showing which UK hackathons are student-run, who ran them, and how big they were, and organiser knowledge gets lost as students graduate.

Demo

Check out the results here!

Code

Check out the folder 0 Weekend here:

Run it with ./run.sh. It sets up the environment, starts Ollama, pulls the model if missing, and writes data/report.md and data/events.csv.

How I Built It

• Model: Gemma 3 (4B), run locally through Ollama.
• Collect: Python collectors read the Hackathons UK season pages and MLH season pages. They then follow each event's own website, plus its About or Team page.
  • Requests respect robots.txt, are rate limited, and are cached.
  • If a site is dead, the pipeline falls back to the Wayback Machine copy from within the same season.
  • Devpost blocks scripted access, so it reads pages you save from a browser.
• Extract: Gemma gets each page and returns JSON matching a schema. It is told to use only what the text states and to leave unknowns as null, with temperature 0 and a verbatim evidence quote for each classification.
• Guardrails: a 4B model guesses student-led from a university name, so code checks the evidence quote.
  • student_led: true needs society or student wording in the quote.
  • false needs company or charity wording.
  • Anything else is demoted to unclear.
• Seasons: seasons run September to August and are named by the ending year (2024 = Sep 2023 to Aug 2024).
• Merge and report: events seen in several sources are merged, with the organiser's own site taking priority. The report is a per-season summary plus a CSV.

Why Does Open Innovation Matter?

• Cost and repeatability: an open-weight model running locally means a pipeline that reads hundreds of pages costs nothing per call and can be re-run whenever the sources change. This is great as it means there is no per-token cost for a volunteer or student organiser who may not have access to LLMs.
• Constrained output: local inference with a JSON schema, temperature 0 and a fixed model version makes extraction reproducible, so anyone can rerun it and check the results.
• Honest limits: because the model is small and open, I had to build the evidence checks. That makes every classification traceable to a quote on a page.

Prize Categories

• Best Use of Gemma
• Best Use of GitHub Copilot