UK Hackathon Statistics

This report details the results of a hackathon project named hackreport. The project is a pipeline designed to read and analyze public pages about past UK hackathons, compiling statistics season by season. The primary focus is on student-led events, recording details such as the organiser, whether the event was student-led, and the size and number of projects involved where available.

The creator built this tool for a colleague who was interested in seeing such statistics across the year. The main issue they aimed to address was the lack of a centralized source showing which UK hackathons are student-run, who organized them, and their scale, as this information often gets lost as students graduate.

The hackreport system is accessible online and the code for the project can be found in the repository named "HacktoberfestOnlineProjects" under AdamDIOM. To run the project, users need to execute "./run.sh" in the folder, which sets up the environment, starts Ollama, pulls the model if necessary, and generates data/report.md and data/events.csv files.

The model used for this project is Gemma 3 (4B), which runs locally through Ollama. The data collection process involves Python scripts that read pages from Hackathons UK season pages and MLH season pages. These scripts follow each event's website and its About or Team pages. Requests to these sites respect the robots.txt file, are rate limited, and are cached.

In case a site is inaccessible, the system falls back to accessing a cached copy from the same season using the Wayback Machine. If a website is blocked by Devpost, the pipeline instead reads pages users have saved from a browser.

The Gemma 3 model processes each page, returning JSON data that matches a predefined schema. It is programmed to use only the information explicitly stated on the page, leaving unknown details as null. The model operates with a temperature of 0 and requires a verbatim evidence quote for each classification it makes. The system has built-in guardrails to ensure accuracy.

For example, when the model guesses an event is student-led based on the presence of a university name, it cross-checks the evidence quote to confirm. The presence of words like "society" or "student" in the quote leads to a student_led classification of true, while references to "company" or "charity" result in a false classification. Any other wording is classified as unclear.

The seasons covered by the project run from September to August, with each season named by the ending year (e.g., 2024 represents events from September 2023 to August 2024). When events are reported in multiple sources, they are merged into a single entry, with preference given to the organiser's own website. The final output includes a per-season summary and a CSV file containing the compiled data.

This project emphasizes the importance of open innovation in terms of cost and repeatability. By running on an open-weight model locally, the pipeline incurs no per-token costs and can be rerun at any time if the source data changes. This makes it particularly useful for volunteer or student organisers who may not have access to large language models.

Additionally, the use of a JSON schema, zero temperature during inference, and a fixed model version ensures that the extraction process is reproducible, allowing anyone to run the pipeline and verify the results themselves. The system also makes its limits clear, with the small and open model necessitating the creation of evidence checks to ensure every classification is traceable to a specific page quote.

The project was submitted in the Hacktoberfest Weekend Challenge, focusing on the "Best Use of Gemma" and "Best Use of GitHub Copilot" categories.