An open Google model on an AMD GPU reads every review of a product, finds the fit and quality problems behind returns, and writes the fix for the listing.
Each report is Gemma reading 60 real reviews of one Rent the Runway item. The grey bar is what renters themselves reported across all reviews of that item, a number Gemma never saw.
Rent the Runway asks every renter whether the item fit, ran small or ran large. We hid that label, gave Gemma only the review text, and compared. The prompt was tuned on separate dev sets. The 40 products in the report above were drawn after the final prompt was written and run once; the full history, including what went wrong, is in the README.
Gemma 3 12B returns fit direction, other causes from a fixed list, the exact quote, and a one-line fix. Batched generation in bf16 on ROCm PyTorch.
The product's "runs small" share and cause mix are plain counts over the per-review answers, so the model cannot inflate a problem.
A second Gemma pass sees only those counts and real quotes, and writes the size note, the description line, and the photo to add.
git clone https://github.com/bisale24-ops/why-they-return && cd why-they-return pip install "transformers>=4.50" accelerate python run_amd.py --set products --out out/products_12b.jsonl # pass 1, AMD GPU python summarize_amd.py --runs out/products_12b.jsonl --out out/product_reports.jsonl python evaluate.py out/products_12b.jsonl # every number on this page