anuj1541/rotary-indictrans2-indic-en-mr-docs

Rotary IndicTrans2 Marathi to English Fine-Tune

Fine-tuned from prajdabre/rotary-indictrans2-indic-en-dist-200M on CIDCO train/val/test + all_documents_translated_progress_1.csv.

Training mode: LoRA (lr=5e-05).

Load

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
repo = "anuj1541/rotary-indictrans2-indic-en-mr-docs"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForSeq2SeqLM.from_pretrained(repo, trust_remote_code=True)

Training

• Direction: mar_Deva -> eng_Latn
• Epochs: 5
• Train pairs: 4,773
• Val pairs: 596
• Effective batch: 32
• LR: 5e-05
• Wall-clock: 2:15:40 (2.26 h)
• Finished: 2026-09-09T22:37:42

Packaged files

pytorch_model.bin, config.json, generation_config.json, tokenization_indictrans.py, tokenizer_config.json, dict.SRC.json, dict.TGT.json, model.SRC, model.TGT, configuration_rotary_indictrans.py, modeling_rotary_indictrans.py

Eval (500 samples, MR→EN, seed=42)

Model	BLEU	ChrF	METEOR	TER	COMET	BERTF1	Judge/5
rotary-base (prajdabre/rotary-indictrans2-indic-en-dist-200M)	39.47	57.76	64.19	76.65	—	92.29	2.80
all-docs (anuj1541/indictrans2-all-docs)	70.56	72.38	74.41	60.22	—	93.58	3.14
fine-tuned (this repo)	49.05	65.92	74.47	73.91	—	93.66	3.20