FineEnvs reports Kannada OCR gains after selecting a checkpoint on test data

FineEnvs says it replaced its test-selected OCR checkpoint with the final training checkpoint on Oct. 13th; current public records identify step 500. The team also reviewed its ASR test curve.

By · Published · Updated

Why it matters

Test-set selection can make a reported score less representative of performance on data the model has not seen. FineEnvs says it changed the OCR adapter after the issue was flagged, but its archive and account do not settle which checkpoint is currently published; the ASR account also says the test curve was reviewed. The results are therefore limited to the named benchmarks and do not independently establish broader performance.

Corrections

  • 2026-10-17: Correction: An earlier version said the run archive still identified OCR step 300 as the selected checkpoint and that the published checkpoint was unclear. Current public documentation identifies step 500 as the published checkpoint, with step 300 retained as a tagged revision. The article also incorrectly assigned a character error rate of 0.3601 to step 500; the current model card assigns 0.3602 to step 500 and 0.3601 to step 300. The passage and summary have been corrected.
A document scanner processes Kannada-script pages at a digitization station, illustrating FineEnvs’ Kannada OCR testing.

FineEnvs reports lower error rates on two Kannada benchmarks: its OCR adapter reduced character error rate from 0.4277 to 0.3602 on 300 Sarvam Indic OCR Bench test crops, while its speech adapter reduced capped per-clip character error rate from 0.1047 to 0.0571 on 838 Kannada FLEURS test clips.…

Reader comments

Conversation for this story loads after sign-in.