⛺ ux-writing-1 · Build Small hackathon

The campfire gallery.

live demo
Judge the blind rewrite battle.

⛺ Copy Campfire

Paste your worst error message; two anonymous campers rewrite it; you vote, then the reveal. Every vote becomes preference data for v2 — the arena is the flywheel.

hf.co/spaces/build-small-hackathon/copy-campfire
the model
Apache-2.0 — yours to run.

🤗 ux-writing-1

27B merged weights, the LoRA adapter, and GGUFs that run on a 24 GB laptop. Model card carries the full honest eval and the tune-it-yourself guide.

hf.co/gr33r/ux-writing-1

The figures behind the video

Small model. Honest numbers.

Trail markers

The four headline stats: blinded preference, throughput, JSON reliability, and what training actually cost.

83% · 7.9K/hr · 99.99% · $30
Real strings. Real rewrites.

Before / after

Unedited rewrites from a 10,000-string review of PostHog, file and line included. Click to deal the next hand of four.

994 changed · 9,006 kept as-is
Same job. Very different bills.

The cost ladder

$3.22 measured on one rented A100 vs the same tokens at frontier list prices — estimates labeled, sources cited.

10,000 strings · 77.2 min · ≈8× cheaper at list
What fine-tuning actually touched.

The LoRA fingerprint

256 measured weight deltas from the released adapter. Hover any cell — the inspector explains what that weight learned, in plain words.

64 layers · late gate_proj burns brightest