hassanql/road-lightonocr2-fit

ROAD LightOnOCR2 fitting pilot

Derivative of LightOn AI's LightOnOCR-2-1B-base, pinned c1bc8eb6625be7b34178ff53481b242f7b8ffcb8. Base model and tokenizer attribution: https://huggingface.co/lightonai/LightOnOCR-2-1B-base Apache License 2.0 is included in LICENSE. Adaptation code and model state are backed up here.

Fixed recipe: four epochs; original grouped fit-only partition; full parameter fine-tuning; language learning rate2e-5 and vision1e-5; batch8 accumulation2; seed42; cosine3%warmup; height128 preserving aspect ratio; Pixtral longest-edge4096; native image-only chat and EOS. Only the predetermined final model is eligible for the frozen validation comparisons. This repository makes no accuracy, rank, or win claim.

Runtime pins: torch2.11.0, torchvision0.26.0, transformers5.14.1, peft0.20.0, accelerate1.14.0. The scripts require original owner-local competition images, CSVs and the frozen split. Those resources, row identifiers, training targets, predictions and private fitting manifests are deliberately absent. Public configuration fingerprints permit owner-local recovery. Checkpoint model/optimizer files permit continuation; final export is inference bf16.

Sources:

• source/train_lighton_ocr.py: preflight, fixed training and validation
• source/lighton_completed_gate_oct2.py: two predefined fixed ensemble gates

Model weights, tokenizer assets, optimizer state and sanitized configuration are the only run artifacts published. Original data must remain inside its licensed competition use.