AKSHARA / TEMPORARY STUDY DASHBOARDTraining and testing progress
OCR studio · Research details ? Final analysis & benchmarks
Windows RTX 3050 · One capped training epoch · Validation selects the model · Final test stays sealed until selection.
ConnectingReading measured run state
Waiting for a progress record.
Frozen training examplesPendingFrozen validation examplesPendingSealed final-test examplesPending
Measured validation results
| Model | Examples | CER | Exact word |
|---|
No candidate improvement is claimed before a supported validation comparison.
Independent final word-crop results
Sealed; no final results are available.
| Model | Examples | Raw CER | NFC CER | Raw WER | NFC WER | Raw exact | NFC exact |
|---|
Output diagnostics, latency, memory and subgroups
These results cannot change validation-locked selection. Intervals are paired by source-document and duplicate components; verified writer identities are unavailable. API parity remains a separate requirement.
Model handoff
Selection has not been locked.
Checking API availability.
Inference pauses during GPU work. Training completion alone does not prove API parity. Final results describe word crops; writer identity, pretraining overlap and complete human label verification remain unknown.
Failure and recovery
No reported coordinator failure.
Checkpoints are versioned and checksum verified. Smoke and benchmark weights are separate from the full training run.