Tahini is sesame: teaching a small open model to spot hidden allergens
说明与用途
BoldPrint 自动生成含有过敏原高亮、营养表和 Nutri‑Score 的标签,将漏检率降至 3%,每千张标签成本 0.12 美元。
问题
Small food sellers need accurate allergen labels but manual labeling is error‑prone and costly.
结果
BoldPrint reduces missed allergens from 74% (untrained model) to 3% (shipped product) on 32 real recipes, achieving $0.12 per thousand labels.
工作流
- Paste recipe text
- Model parses ingredients and identifies allergens
- Cross‑check with Open Food Facts lexicon
- Generate nutrition table and Nutri‑Score
- Export PNG or PDF sticker
- Optional human review of ⚠ Check lines
证据
On 32 real-world recipes I wrote and labelled by hand (brands, typos, several languages): ... BoldPrint as shipped (model + lexicon cross-check) 3.0% missed allergens, 2/32 labels with a miss.
The lexicon alone missed 11% of the allergens, and 9 of the 32 labels had at least one miss.
Over 20 RL steps, missed allergens per label on the training batches fell from about 0.05 to under 0.02, and the average reward roughly tripled (0.22 → 0.66).
- 记录日期: 2026-10-17T10:41:20.647Z