Note Environment · document OCR, layout detection and VQA over 1M Nayana pages in 22 languages, plus Sarvam Indic OCR Bench as evaluation splits. Playground at /web.

Note Environment · transcription, verbatim transcription and language ID over all 102 FLEURS languages, scored per word or per character by script.

FineEnvs/gemma-4-E4B-it-kannada-ocr-grpo

Image-Text-to-Text • Updated 1 day ago • 19

Note Kannada OCR · Gemma 4 E4B + LoRA, GRPO on 4,000 Nayana crops. Sarvam Indic OCR Bench (Kannada) CER 0.428 → 0.360.

FineEnvs/gemma-4-E4B-it-kannada-asr-grpo

Automatic Speech Recognition • Updated 1 day ago • 15

Note Kannada ASR · Gemma 4 E4B + LoRA, GRPO on every FLEURS Kannada clip. Held-out CER 0.105 → 0.057 (−45%).

Note Training and live-evaluation curves for both Kannada runs.

Note Every checkpoint's held-out predictions, paired curves, trainer logs and the exact HF Jobs / TRL scripts that ran.

Note Sarvam Indic OCR Bench by Sarvam AI (Apache-2.0), served as evaluation-only splits and scored with its own metrics.py.

Note Source corpus for the OCR environment, by CognitiveLab (CC BY-NC 4.0).

Note Source speech for the ASR environment, by Google (CC BY 4.0).

google/gemma-4-E4B-it

Any-to-Any • 8B • Updated Jul 20 • 4.42M • 1.64k

Note Base model for both adapters: Gemma 4 E4B, with audio and vision encoders.