LightOnOCR-3 Training Data Revealed: MinHash Dedup, Weighted Formula/Table Sampling, Muon
IgorCarron · x · 2026-10-09
LightOn details LightOnOCR-3's training data recipe: MinHash deduplication across the full corpus, 2× sampling for formulas and 3× for tables, handwriting and multilingual content, synthetic documents via in-house Docduck, better filtering with extensive image augmentation, and training with the Muon optimizer.
More from Models
- Open-source models keep trailing the frontier — for now we can pick our ideology — panickssery · 2026-10-09
- OpenAI rolls out Ultrafast mode for GPT-6.1 Sol at up to 8x standard speed — OpenAIDevs · 2026-10-09
- Dev says GPT-6.1 Sol one-shot full controller support for his game — Dimillian · 2026-10-09
- Dev Says GPT-6.1 Sol One-Shotted Full Gamepad Support for His Game — Dimillian · 2026-10-09
- Haiku 5.5 Hits 85% Success on Robot Tasks at Under $0.02 Per Attempt — scaling01 · 2026-10-09
- LangChain Podcast Digs Into Decision Models as OpenAI and Databricks Enter the Space — LangChain · 2026-10-09