Modelpublic

build-small-hackathon/ux-writing-1-lora

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

1likes12downloads

Model Card

ux-writing-1-lora

🏕️ Hackathon mirror. Canonical home: `gr33r/ux-writing-1-lora` · merged model: `gr33r/ux-writing-1`

The QLoRA adapter behind **`gr33r/ux-writing-1`** — an open UX writing reviewer that takes a UI string + its code context and returns compact JSON {"rewrite", "reason", "risk"}. Blind-validated at 83% human preference over the fair-prompted base model (65/78 decisive comparisons on 90 held-out items — methodology and eval code in the repo).

Built for the Hugging Face Build Small hackathon — try it in the ⛺ Copy Campfire arena.

Which artifact do you want?

you want	use
Just run the model	`gr33r/ux-writing-1` (merged, vanilla transformers)
Run it on a laptop	`gr33r/ux-writing-1-GGUF` (Q4KM 16.6 GB)
This repo: attach to base / continue training	the 159 MB LoRA (r=16 α=32, LM projections)

Usage (PEFT)

python

import torch
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoTokenizer

base = "Qwen/Qwen3.6-27B"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForImageTextToText.from_pretrained(base, dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, "gr33r/ux-writing-1-lora")
# Prompt contract + enable_thinking=False guidance: see the merged model card.

Training: QLoRA (4-bit NF4, double-quant, bf16 compute), LoRA on q,k,v,o,gate,up,down projections, 2 epochs on ≈1,400 owner-authored/derived rewrite pairs, one A100-80GB. To fine-tune further on your style guide (≈$2–6 on HF Jobs), see FINETUNE_GUIDE.md.

License: Apache-2.0. Attribution appreciated: ux-writing-1 by gr33r.