Downloads · 30 days
0
Arushhh/indic-translate-mr-lora
indic-translate-mr-lora is a translation model from Arushhh. Use it when you need text moved from one language to another. It is set up for peft. The card lists the license as other.
LoRA adapters and evaluation artifacts from fine-tuning bodhan-ai/indic-translate on English–Marathi data.
Downloads · 30 days
0
Access
Public
Updated Sep 23, 2026
Repo size
—
Likes
0
Public
Click a slice to open those files.
.jsonl3.4 MB · 100%
From the Hugging Face model README
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.
| 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.
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.
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.