1 Training data← Back to the data viewer2 TokenizerCompareAuditExplainer slidesCoordinate tokens3 Image budget4 MP vs 8 MP4 NotesData processingAssamese ra checkSynthetic checkSmall held-out blocksData EDAJaigyan latticeJaigyan v4 re-readEmpty pagesManipuri: 4B vs pipeline5 TrainingEmbedding warm-upSFT ablationsSFT merges2B distillation0.8B distillationDistillation explainedRL planRL ablationsRL training

Embedding warm-up

Connecting to the inference service…

Progress: CER per checkpoint

Greedy decoding of the full validation set: 50 crops per language (Urdu from gemini_urdu, Kashmiri from synth_kashmiri) + 50 arXiv pages (arXiv CER on text only, location tokens removed). CER per sample is capped at 100%, so one loop cannot dominate a mean; "Loops" = share of outputs that reach the token limit. Stock = the original Qwen3.5-4B with its own tokenizer; step 0 = our model before training (initialized new rows). Output tokens count with each model's own tokenizer. New checkpoints are read automatically. Lower is better.

Try a checkpoint

Selected image

This sample over training

Predictions saved by the CER check, at each evaluated checkpoint.