build-small-hackathon/hackathon-advisor

1---2license: apache-2.03task_categories:4- text-classification5- text-generation6language:7- en8tags:9- hackathon-advisor10- quest-classification11- lora-sft12- minicpm513pretty_name: Hackathon Advisor Quest Classification SFT14size_categories:15- n<1K16---17 18# Hackathon Advisor — Quest Classification SFT Dataset19 20Supervised fine-tuning data that teaches MiniCPM5-1B to classify a Build Small21Hackathon project against 13 judging dimensions from a two-segment README + app-file22prompt, emitting strict JSON with short, source-attributed evidence. Trains the LoRA at23[`build-small-hackathon/hackathon-advisor-quest-minicpm5-lora`](https://huggingface.co/build-small-hackathon/hackathon-advisor-quest-minicpm5-lora).24 25## Format (`quest_sft.jsonl`)26 27Chat-JSONL. The **first line** is a `lora_sft_manifest`; every following line is a28`lora_sft_example` with a `messages` list (system / user / assistant). The assistant29turn is exactly one JSON object:30 31```json32{"matches":[{"quest":"...","confidence":0.0,"evidence":"...","source":"readme|app_file"}]}33```34 35No markdown, no prose, no renamed quests; an empty `matches` list when no dimension has36clear evidence. The user turn splits the project into a `[README]` segment and an37`[APP_FILE]` segment so the model judges product description and implementation38evidence separately and attributes each match to its source.39 40## Quest dimensions (13)41 42Six merit badges (Off the Grid, Well-Tuned, Off-Brand, Llama Champion, Sharing is43Caring, Field Notes), two tracks (Backyard AI, Thousand Token Wood), and five44sponsor / special awards (OpenBMB, Nemotron, Modal, Tiny Titan, Best Agent).45 46## Examples: 156 (14 with empty matches)47 48| variant | count |49| --- | --- |50| natural | 108 |51| app_only | 16 |52| missing_app_file | 16 |53| noisy_metadata | 8 |54| contradiction | 6 |55| empty | 2 |56 57Positive examples per quest:58 59| quest | examples |60| --- | --- |61| Off the Grid | 87 |62| Off-Brand | 59 |63| Tiny Titan | 58 |64| Thousand Token Wood | 49 |65| Llama Champion | 35 |66| Backyard AI | 35 |67| Well-Tuned | 31 |68| OpenBMB | 26 |69| Sharing is Caring | 19 |70| Nemotron | 18 |71| Field Notes | 15 |72| Modal | 14 |73| Best Agent | 14 |74 75## Provenance76 77Built from the real public Spaces of the `build-small-hackathon` org: 125 crawled78projects → deduped + length-filtered to 108 content-rich ones → labelled by a79teacher-then-adversarial-verifier multi-agent workflow → plus targeted augmentations80(app-only, readme-only / missing app file, README↔app contradictions, empty matches,81noisy metadata). `labeled.json` holds the per-project verified labels. Examples are82derived from public hackathon submissions for research and hackathon use; each project83remains under its own Space license.84