indic-translate Marathi LoRA — experiment artifacts
LoRA adapters and evaluation artifacts from fine-tuning
bodhan-ai/indic-translate
on English–Marathi data.
This is a derivative work of bodhan-ai/indic-translate (Bodhan AI / AI4Bharat),
used and redistributed under the Indic Open Model License v1.0. The adapters
here are not usable alone; they require the base model at load time.
What is here
| path | contents |
|---|---|
generations/ |
raw model outputs per direction — the primary artifact |
metrics/ |
chrF++ (word_order=2) + BLEU with bootstrap CIs and sacrebleu signatures |
checkpoint-*/ |
LoRA adapters, pushed during training |
Raw generations are kept deliberately: metrics can be recomputed from generations, but generations cannot be recovered from metrics.
Experiment
Three matched 25k-pair training arms isolating two axes:
| arm | corpus | quality | domain |
|---|---|---|---|
| A | Samanantar (structural filters only) | noisy | general/news/govt |
| B | BPCC_cleaned (LaBSE ≥ 0.9) | clean | general/news/govt |
| C | shiksha (NPTEL lectures) | clean-ish | education |
Training data was de-duplicated against IN22-Gen before use.
Caveat on metrics
chrF++ differences below roughly ±0.7 are inside the test-set bootstrap CI at n=1024 and should not be read as real. The ±0.06 figure quoted in the toolkit is decode-repeat noise, a different and much smaller quantity.
- Downloads last month
- -