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 · Data quality audit

Inspect the evidence before changing supervision.

Automatic flags are review candidates. Repetition can be legitimate in tables, exercises and forms. Unflagged does not mean correct: fluent invented text, substitutions and omissions can pass these checks. Counts below describe saved OCR outputs, not confirmed hallucinations.

Replacement OCR · best_ckpt_rl · node 40

Loading redo status…

Distribution · nonempty blocks

Detector evidence · rules can overlap

Loading…
Select a block to inspect.
Method, provenance and limits

Language comes from the existing source index; script mismatch is advisory. Text density uses source crop pixels and Unicode code points. No OCR model runs here. Image geometry is checked against original annotations on selection. Unflagged controls are a deterministic sample across source, language and length, not an accuracy evaluation sample. Review notes are append-only and do not modify dataset annotations or OCR sidecars.

Full source page with selected block highlighted