FineEnvs/multilingual-multimodal-rl-runs

Multilingual Multimodal RL runs

The evidence behind the two Kannada GRPO runs in the Multilingual Multimodal Envs collection: what the model wrote for every held-out item at every checkpoint, the curves, the training logs, and the exact scripts that ran. If a number on one of the model cards looks wrong, this is where to check it.

Run	Model	Trained on	Held-out set	Base → selected checkpoint
asr-kannada	gemma-4-E4B-it-kannada-asr-grpo	all 2,282 FLEURS kn_in train clips	FLEURS kn_in test, 838 clips	CER 0.1047 → 0.0571 (step 575)
ocr-kannada	gemma-4-E4B-it-kannada-ocr-grpo	4,000 Nayana Kannada section crops	Sarvam Indic OCR Bench test, Kannada, 300 crops	Sarvam CER 0.4277 → 0.3601 (step 300)

Both runs fine-tune google/gemma-4-E4B-it with LoRA and GRPO from TRL. The reward comes from an OpenEnv server: 0.8 × (1 − character error) + 0.2 × exact match. While each run trained, a second job scored every checkpoint on held-out data as soon as it was saved, with greedy vLLM decoding, graded by the same server.

Tables

Config	Rows	One row per
curves	45	run × scored checkpoint: means, and the paired change from base with a 95% interval
asr_kannada_evals	20,112	checkpoint × test clip: reference, prediction, reward, CER, WER
ocr_kannada_evals	6,300	checkpoint × benchmark crop: reference, prediction, reward, CER, Sarvam CER/WER, loop flag

from datasets import load_dataset

curves = load_dataset("FineEnvs/multilingual-multimodal-rl-runs", "curves", split="train")
asr = load_dataset("FineEnvs/multilingual-multimodal-rl-runs", "asr_kannada_evals", split="train")
final = asr.filter(lambda r: r["checkpoint"] == "step-575")

Error rates in curves are capped at 1 per item before averaging, so one looping prediction cannot dominate a checkpoint. Sarvam's metric already caps them. The per-item tables keep the raw values. step 0 is the untuned base model.

Files

asr-kannada/ and ocr-kannada/
├── run.json                  model, selected checkpoint, commit, job ids, both launch commands
├── summary.json              the trainer's own end-of-run check
├── run-metadata.json         resolved config, LoRA targets, train task ids, eval set id
├── baseline.json             the in-training eval at step 0 (48 items, transformers generate)
├── trainer_log_history.jsonl every logged training step: reward, loss, grad norm, lengths, ...
├── evals/                    base.json and step-N.json: every held-out prediction, per checkpoint
├── plots/                    curves.gif, process.gif, eval_curve.png, eval_change.png, training.png
└── code/                     the job launcher, trainer, evaluator and reward at the commit that ran

The code snapshots are verbatim from the run commits (ac63930 for ASR, 6e8dd18 for OCR). The folders have moved since the runs, from 05-multilingual-ocr and 06-multilingual-asr to the two halves of one project, 06-multilingual/ocr and 06-multilingual/asr.

Sources and licenses

ASR references are FLEURS transcriptions (google/fleurs, CC BY 4.0). OCR references are from Sarvam Indic OCR Bench by Sarvam AI (Apache-2.0); its metrics.py produced the Sarvam CER/WER columns. Predictions and code are released under the licenses of their sources and of FineEnvs (Apache-2.0).

@misc{fineenvs,
  author = {Kolavi, Adithya S},
  title  = {FineEnvs: Open Source RL Environments for LLM Agents},
  year   = {2026},
  url    = {https://github.com/adithya-s-k/FineEnvs}
}