{"id":50700271,"url":"https://github.com/twangodev/devpost-hacks","last_synced_at":"2026-06-09T09:02:03.508Z","repository":{"id":354584487,"uuid":"1218582442","full_name":"twangodev/devpost-hacks","owner":"twangodev","description":"Can LLMs judge hackathons?","archived":false,"fork":false,"pushed_at":"2026-05-05T01:58:34.000Z","size":124,"stargazers_count":0,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-05-05T03:37:53.933Z","etag":null,"topics":[],"latest_commit_sha":null,"homepage":"","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":null,"status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/twangodev.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":null,"code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-04-23T02:41:02.000Z","updated_at":"2026-05-05T01:58:38.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/twangodev/devpost-hacks","commit_stats":null,"previous_names":["twangodev/eval-hack","twangodev/devpost-hacks"],"tags_count":null,"template":false,"template_full_name":null,"purl":"pkg:github/twangodev/devpost-hacks","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/twangodev%2Fdevpost-hacks","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/twangodev%2Fdevpost-hacks/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/twangodev%2Fdevpost-hacks/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/twangodev%2Fdevpost-hacks/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/twangodev","download_url":"https://codeload.github.com/twangodev/devpost-hacks/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/twangodev%2Fdevpost-hacks/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":34098952,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-06-09T02:00:06.510Z","response_time":63,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":[],"created_at":"2026-06-09T09:02:00.742Z","updated_at":"2026-06-09T09:02:03.492Z","avatar_url":"https://github.com/twangodev.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"# Can LLMs Judge Hackathons?\n\n[![HF Collection](https://img.shields.io/badge/Hugging%20Face-Collection-222?logo=huggingface\u0026logoColor=FFD21E)](https://huggingface.co/collections/twangodev/devpost-hacks)\n[![Report](https://img.shields.io/badge/Report-PDF-b31b1b)](report.pdf)\n\nEvaluating whether LLMs can judge hackathons in agreement with human judges.\nScraping lives in a separate repo; this one covers enrichment, pairwise\njudging, Bradley-Terry ranking, and dataset publishing.\n\nWe compare five judges:\n\n1. **Qwen3.5-27B** — large reasoning teacher\n2. **Qwen3.5-4B** — small open-weights reasoning peer\n3. **Qwen3-4B-Instruct-2507** — the actual base under our LoRA\n4. **twangodev/devpost-hacks-qwen3-4b-judge** — our 4B finetune distilled from the 27B teacher\n5. **gpt-oss-20b** — cross-family open-weights baseline\n\nPairwise verdicts feed a Bradley-Terry model. Primary eval: top-1 agreement\nwith human winners on Devpost. Secondary: percentile correlation against the\nprivate MadHacks Plackett-Luce ranking.\n\n## Datasets\n\n- [`twangodev/devpost-hacks`](https://huggingface.co/datasets/twangodev/devpost-hacks)\n  — 2,222 projects across 9 hackathons + GitHub READMEs\n- [`twangodev/devpost-hacks-judgments`](https://huggingface.co/datasets/twangodev/devpost-hacks-judgments)\n  — SFT-format pairwise judgments from each of the five judges above; row's\n  `model` field disambiguates the source\n\n## Pipeline\n\n```bash\nuv sync\ndevpost enrich -s \u003chackathon\u003e.json             # GitHub READMEs → projects parquet\ndevpost pairs  -c \u003chackathon\u003e --n 500          # sample pairs from the HF dataset\ndocker compose up -d sglang-qwen3.5-27b        # one of: sglang-qwen3.5-{27b,4b}, sglang-qwen3-4b-{2507,judge}, sglang-gpt-oss-20b\ndevpost judge  --pairs data/\u003ch\u003e/pairs.jsonl --model Qwen/Qwen3.5-27B    # resumable; matches the running service\ndevpost rank   -c \u003chackathon\u003e                  # BT ranking + recall@K to stdout\ndevpost summary                                # cross-judge eval w/ bootstrap CIs and pairwise stats\ndevpost export projects | judgments            # stage for HF upload\n```\n\nExternal: `gh` (README fetches), Docker (SGLang).\n\n## Authors\n\nManas Joshi · Oliver Ohrt · Andrew Lou · Mari Garey · George Sukhotin · Ethan Feiges · Ben Friedl · James Ding\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Ftwangodev%2Fdevpost-hacks","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Ftwangodev%2Fdevpost-hacks","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Ftwangodev%2Fdevpost-hacks/lists"}