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
Unified OCR · Qwen3.5-4B

SFT merges

About this page

Weight merges of SFT checkpoints (merged-ckpts/sft/, one folder per merge with MERGE.json). Each merge is scored on the same 4 benchmarks as the training checkpoints: OmniDocBench v1.6 EN, olmOCR-Bench EN, Indic printed and Indic handwritten crops (word accuracy = 100 × (1 − house WER clean), as the blog). Member rows (grey) are the checkpoints that went into the merges. The plan, the pass bar and the findings are below the tables. Refreshes every minute.

Loading…

Merges vs their members green = best per column; Δ = vs the best member on that column; verdict = Omni and both Indic WERs strictly better than the best member; olmOCR ≈ 83.7 is enough (user)

Indic per language WER (clean), lower is better; green = best per column