README.md

metadata

title: Ready to Submit?
emoji: 👁
colorFrom: gray
colorTo: green
sdk: gradio
sdk_version: 6.18.0
python_version: '3.12'
app_file: app.py
startup_duration_timeout: 45min
pinned: false
license: mit
short_description: Evaluates your HF Space for Build Small Hackathon
tags:
  - track:backyard
  - sponsor:nvidia
  - sponsor:openbmb
  - achievement:offbrand
models:
  - nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16
  - JetBrains/Mellum2-12B-A2.5B-Instruct
  - openbmb/MiniCPM5-1B

👁 Ready to Submit?

The app is the question. Point it at any Space in the build-small-hackathon org and it checks the entry rules from the official field guide — then a small model (your pick, under 32B of course) writes you a grounded, actionable review.

The idea

The hackathon has six entry rules, two tracks, four sponsor prizes, six achievement badges and six judged bonus awards — and the difference between "submitted" and "eligible" hides in README frontmatter tags like track:backyard and achievement:offgrid. Ready to Submit? automates the pre-flight check: it verifies the verifiable (deterministically, via the HF Hub API) and lets a small model handle the judgment calls (track fit, prize opportunities, README polish), grounded in the machine-verified facts so it can't make things up.

How it works

1. Grounded checks (no LLM): fetches the target Space's metadata, README and source via the public Hub API; parses the frontmatter tags against the canonical ids from the field guide's own source; finds demo-video and social-post links; detects every Hub model referenced by the app and looks up its real parameter count against the 32B cap (and the ≤4B Tiny Titan bar).
2. Small-model review: the checklist + facts + rules digest go to the reviewer model you picked, which streams back fixes, track-fit reasoning, and the prizes/badges the Space could claim but hasn't.

Tech

• Models (pick your reviewer): NVIDIA Nemotron 3 Nano 4B (default — 3.97B params, a hybrid Mamba-Transformer that even fits the Tiny Titan bar), JetBrains Mellum 2 12B-A2.5B Instruct, OpenBMB MiniCPM5 1B.
• Runtime: gr.Server() on ZeroGPU — plain FastAPI routes serve a custom HTMX frontend (no stock Gradio components anywhere), and the review streams through a Gradio-queued endpoint via @gradio/client, transformers + bf16, TextIteratorStreamer.
• Custom UI: hand-rolled pastel re-skin of the field guide's woodblock-press design language — paper grain, dashed-ring stamp badges, self-hosted Archivo/Spline Sans Mono, htmx swaps with a friendly loading stamp. That's the achievement:offbrand story.
• Grounding: rules and canonical tags extracted from the field guide Space's source of truth (src/lib/data/content.ts, src/lib/readme.ts), embedded as the reviewer's system context; checks (including Codex commit attribution and per-model parameter counts) are deterministic Hub API calls, so the model can't invent facts.
• Honesty: every AI review ships with a disclaimer — double-check against the official field guide regardless of what the app says.

Links

• 🎬 Demo video: demo.mp4
• 📣 Social post: https://x.com/amphetamarina/status/2065435918509441045