"""Field-guide digest used to ground the LLM recommendations. Sourced from the official Build Small field guide (https://build-small-hackathon-field-guide.hf.space/) — rules, tags and prize data extracted from the guide's own source of truth (`src/lib/data/content.ts` and `src/lib/readme.ts` in build-small-hackathon/field-guide). """ ORG = "build-small-hackathon" DEADLINE = "June 15, 2026" TRACK_TAGS = { "track:backyard": "Backyard AI — practical, problem-solving apps that improve daily life", "track:wood": "Thousand Token Wood — whimsical, delightful, AI-native fun", } SPONSOR_TAGS = { "sponsor:openbmb": "OpenBMB prize — requires building with MiniCPM models", "sponsor:openai": "OpenAI prize — requires Codex-attributed commits in the repo/space", "sponsor:nvidia": "NVIDIA prize — requires using Nemotron models", "sponsor:modal": "Modal prize — requires Modal platform for dev or runtime, noted in README", } ACHIEVEMENT_TAGS = { "achievement:offgrid": "Off the Grid — no cloud APIs, everything runs on local models", "achievement:welltuned": "Well-Tuned — uses a fine-tuned model you published on the Hub", "achievement:offbrand": "Off-Brand — custom frontend pushing past the default Gradio look", "achievement:llama": "Llama Champion — model runs through the llama.cpp runtime", "achievement:sharing": "Sharing is Caring — agent trace shared on the Hub", "achievement:fieldnotes": "Field Notes — blog post or report about the build", } MANAGED_PREFIXES = ("track:", "sponsor:", "achievement:") # Lucide glyph per tag — same mapping the field guide site uses (Icon.svelte) TAG_GLYPHS = { "track:backyard": "tent", "track:wood": "trees", "achievement:offgrid": "unplug", "achievement:welltuned": "target", "achievement:offbrand": "palette", "achievement:llama": "cpu", "achievement:sharing": "share-2", "achievement:fieldnotes": "notebook-pen", "sponsor:openbmb": "sparkles", "sponsor:openai": "circle-play", "sponsor:nvidia": "cpu", "sponsor:modal": "mountain", } TAG_SHORT_NAMES = { "track:backyard": "Backyard AI", "track:wood": "Thousand Token Wood", "achievement:offgrid": "Off the Grid", "achievement:welltuned": "Well-Tuned", "achievement:offbrand": "Off-Brand", "achievement:llama": "Llama Champion", "achievement:sharing": "Sharing is Caring", "achievement:fieldnotes": "Field Notes", "sponsor:openbmb": "OpenBMB", "sponsor:openai": "OpenAI · Codex", "sponsor:nvidia": "NVIDIA", "sponsor:modal": "Modal", } # Compact cheat-sheet for the reviewer prompt — small (1B) models echo long # briefings back, so the LLM gets this instead of the full digest below. GUIDE_BRIEF = """\ Build Small hackathon, deadline June 15 2026. Six entry rules: every model under 32B params; Gradio app in the build-small-hackathon org; demo video linked in README; social post linked in README; max 10 Zero GPU apps; README frontmatter tags + short write-up. Tracks (tag at least one): `track:backyard` practical daily-life apps, `track:wood` whimsical fun apps. Track prizes $4k/$2.5k/$1.5k/$1k + $2k community choice. Sponsor prize tags: `sponsor:openbmb` needs MiniCPM models; `sponsor:openai` needs Codex-attributed commits; `sponsor:nvidia` needs Nemotron models; `sponsor:modal` needs Modal usage noted in README. Achievement tags (no prize): `achievement:offgrid` local only, no cloud APIs; `achievement:welltuned` own published fine-tune; `achievement:offbrand` custom UI; `achievement:llama` llama.cpp; `achievement:sharing` trace shared on Hub; `achievement:fieldnotes` blog/report. Judged cash bonuses (not taggable): Off Brand $1.5k custom UI; Tiny Titan $1.5k best <=4B-model app; Best Demo $1k; Best Agent $1k; Bonus Quest Champion $2k; Judges' Wildcard $1k. """ GUIDE_DIGEST = """\ # Build Small Hackathon — official rules digest Hugging Face x Gradio hackathon for small (<32B parameter) models. Final submission deadline: June 15, 2026. Submissions live as Spaces in the `build-small-hackathon` org on Hugging Face. ## The six entry rules (ALL required to submit) 1. Stay under 32B — every model used must individually have fewer than 32B total parameters (not just active). Combining several small models is fine. 2. Ship a Gradio app — deploy as a Gradio App inside the official org. Docker Spaces are fine as long as the interface is a Gradio Space. 3. Record a demo — a demo video showing the app working, publicly watchable (YouTube, a file in the Space, or any public host), linked from the README. 4. Post it — one social-media post showcasing the app, linked from the README. 5. Mind the GPU limit — max 10 Zero GPU apps per user when using provided Zero GPU resources. 6. Tag your README — frontmatter `tags:` must declare the tracks/badges to be considered for, plus a short write-up of the idea and the tech used. ## Tracks (tag at least one) - `track:backyard` — Backyard AI: practical, problem-solving apps for daily life (storybook generator for a child, study tutor, receipt parser). - `track:wood` — Thousand Token Wood: whimsical, delightful, AI-native fun (games, desktop pets, text-adventure dungeon masters). Prizes per track: $4k / $2.5k / $1.5k / $1k + $2k community choice. ## Sponsor prizes (tag to enter; each has a hard requirement) - `sponsor:openbmb` — must build with MiniCPM models ($2.5k/$1.5k/$1k per track pool). - `sponsor:openai` — requires Codex-attributed commits; holistic Codex use ranks higher ($5k/$3k/$1k). - `sponsor:nvidia` — must use Nemotron models (2x RTX 5080, best space + community engagement). - `sponsor:modal` — must use Modal for development or runtime and note it in the README ($10k/$7k/$3k in credits). ## Achievement badges (no prize, self-tagged bragging rights) - `achievement:offgrid` — no cloud APIs at all, local models only. - `achievement:welltuned` — uses a fine-tuned model published on the Hub. - `achievement:offbrand` — custom UI well past the default Gradio look. - `achievement:llama` — model runs through llama.cpp. - `achievement:sharing` — agent trace shared on the Hub. - `achievement:fieldnotes` — blog post / report about the build. ## Cash bonus awards (judged — cannot be self-tagged) - Off Brand ($1.5k): best custom UI past the default Gradio look. - Tiny Titan ($1.5k): best app on a genuinely tiny model, <=4B parameters. - Best Demo ($1k): great app + great video + great social post. - Best Agent ($1k): best agentic app under the 32B cap. - Bonus Quest Champion ($2k): most bonus criteria met; ties go to ambition + quality. - Judges' Wildcard ($1k): amazing but fits no category; every entry is in the running. ## Submission workflow Join the org -> upload the Space -> record + host the demo video -> social post -> README updated with both links, the frontmatter tags, and a short write-up (idea, how it was built, tech used). One project can stack a track placement, several sponsor prizes and bonus awards. Multiple submissions are allowed; each Space is judged independently. """