| --- |
| title: Copy Campfire |
| emoji: ⛺ |
| colorFrom: yellow |
| colorTo: red |
| sdk: gradio |
| sdk_version: "5.49.1" |
| app_file: app.py |
| pinned: false |
| license: apache-2.0 |
| short_description: UX writing arena — base model vs fine-tune. You judge. |
| tags: |
| - track:backyard |
| - sponsor:modal |
| - achievement:welltuned |
| - achievement:offbrand |
| - achievement:fieldnotes |
| --- |
| |
| # ⛺ Copy Campfire — the UX writing arena |
|
|
| Two UX writers by the fire — one went to training camp. Paste your UI copy, get two |
| blind rewrites (base **Qwen3.6-27B** vs the **ux-writing-1** QLoRA fine-tune), vote, |
| then see the reveal. Votes are stored privately as preference data that trains the next |
| version — the arena is the data flywheel. |
|
|
| ## The idea |
|
|
| Every product ships copy debt — "Invalid", "OK", "An error occurred while processing |
| your request" — thousands of interface strings nobody owns. **ux-writing-1** is a small, |
| open model fine-tuned to review that copy the way a senior UX writer would: rewrite |
| what's weak, **keep what's already right**, and say why. It runs at codebase scale for |
| cents, on hardware you control — exactly the kind of bulk, structured LLM work that |
| shouldn't be metered at frontier prices. |
|
|
| ## How it was built |
|
|
| - **Base:** [Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B) — Apache-2.0, 27.8B |
| params (under the 32B limit). |
| - **Training:** QLoRA on ~1,400 hand-built before/after pairs, each with a |
| `{rewrite, reason, risk}` contract and the code context the string lives in. Two runs, |
| ≈$30 of GPU on **Modal**. |
| - **Serving:** Modal OpenAI-compatible endpoint; this Gradio arena calls it live. |
| - **Honest eval:** 83% blinded human preference vs the base model (90 held-out items) — |
| not automated heuristics, which saturated. |
| - **At codebase scale:** reviewed 10,000 of PostHog's UI strings in 77 min for **$3.22** |
| — changed 994, **kept 9,006** as-is. Restraint is trained, not hoped for. |
|
|
| ## Tech |
|
|
| Qwen3.6-27B · QLoRA (PEFT / TRL) · Modal (training + serving) · Gradio (this arena) · |
| llama.cpp / GGUF (zero-cost laptop inference) · Hugging Face Hub (model, adapter, |
| dataset, blinded vote store). |
|
|
| ## Take it home |
|
|
| - **Model:** [`gr33r/ux-writing-1`](https://huggingface.co/gr33r/ux-writing-1) · |
| [adapter](https://huggingface.co/gr33r/ux-writing-1-lora) · |
| [GGUF](https://huggingface.co/gr33r/ux-writing-1-GGUF) |
| - **Code & eval tooling:** [`content-designer/ux-writing-1`](https://github.com/content-designer/ux-writing-1) |
| - **Interactive figures:** [the campfire gallery](https://copy-campfire-gallery.vercel.app) |
|
|
| ## Submission |
|
|
| - 🎬 **Demo video:** [Watch the 2-minute demo](https://www.linkedin.com/posts/christopher-greer_buildsmall-opensource-ai-ugcPost-7472300204092764161-SnJ9/) — also embedded in the **🎬 Watch the demo** tab inside the Space. |
| - 📣 **Social post:** [On LinkedIn](https://www.linkedin.com/posts/christopher-greer_buildsmall-opensource-ai-ugcPost-7472300204092764161-SnJ9/) |
|
|
| --- |
|
|
| **Space secrets required:** `BATTLE_URL`, `AUTH_TOKEN` (Modal backend), `HF_TOKEN` |
| (write access for vote storage); optional `VOTES_DATASET`. |
|
|
