Datalab details the OCR script confusion bug and cuts it 7x with a new training technique
VikParuchuri · x · 2026-10-16
Vik Paruchuri's thread wraps up Datalab's full write-up on "script confusions" in OCR models:
- The bug: correct word, wrong alphabet — "vain" became 罪 (Chinese for sin) on a hymnal page; Finnish LEPPÄVIRRAN ended in Arabic ران; Tarjoamme ended in Greek ουμεν
- Rare (a fraction of a percent of pages) but sneaky: edit-distance metrics barely move and text search won't find the error
- Frontier models affected too, including Gemini 3.8 Flash and Claude Opus 5.5
- Likely cause: same-meaning words across scripts share nearly identical embeddings inside the model
- Fix: Chandra 2.5, trained with a novel technique, shows 7x fewer script confusions than Chandra OCR 2.4 on a hard test set, outperforming both frontier models; live on the API now
More from Models
- NEURA launches Neuron AI, a multimodal PLA model giving robots touch, force feedback and memory — CyberRobooo · 2026-10-17
- Sentience claims per-user models: train, continually update and serve one model per person — soleio · 2026-10-17
- Four Gemini 4 Argon agents beat Minecraft in 27 hours, writing 2,783 programs — GoogleDeepMind · 2026-10-17
- Meta open-sources Muse Spark Open to mixed reception: 'trash' license, nerfed benchmark scores — gabriberton · 2026-10-17
- If new diffusion models are open-weight only, what's left for commercial use? — blackpanda8484 · 2026-10-17
- Reflection Beam OSS architecture: sliding window attention, gated attention, QK Norm — A_K_Nain · 2026-10-17