Hackathon Advisor — Quest Classification LoRA (MiniCPM5-1B)

PEFT LoRA adapter that classifies a Build Small Hackathon project against 13 judging dimensions (6 merit badges + 2 tracks + 5 sponsor/special awards) from a two-segment README + app-file prompt, emitting strict JSON:

{"matches":[{"quest":"...","confidence":0.0,"evidence":"...","source":"readme|app_file"}]}

Load it in the deployed Space by setting ADVISOR_QUEST_ADAPTER_ID to this repo. The backend revalidates every dashboard refresh and will not swap on schema failure.

Recipe

• Base model: openbmb/MiniCPM5-1B
• Task: hackathon_advisor_quest_classification
• Method: LoRA SFT (completion-only loss)
• Examples: 259
• Epochs: 16.0
• LoRA rank/alpha/dropout: 64/128/0.0
• Max seq length: 3072
• GPU: L40S

Dataset

build-small-hackathon/hackathon-advisor-quest-dataset — 156 chat-JSONL examples built from real build-small-hackathon Spaces: 108 teacher- labelled + adversarially-verified projects plus targeted augmentations (app-only, readme-only / missing app file, README↔app contradictions, empty matches, noisy metadata). All 13 quests covered.

Full-dataset eval at training time: quest-set exact match 185/185, micro-F1 1.0.

Evaluated by reproducing the gold quest set for every example in the training dataset (the dataset is the spec — it is built from the real build-small-hackathon projects).