IndicTrans2 English to Marathi Fine-Tune

Fine-tuned from ai4bharat/indictrans2-en-indic-dist-200M on the combined anuj1541/cidco-marathi-english-qa-scored and anuj1541/marathi-english-cidco-long-docs datasets.

Training mode: LoRA merged into the base (r=16, lr=0.0001).

Load

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

Training

  • Direction: eng_Latn -> mar_Deva
  • Epochs: 3
  • Train pairs: 58,761
  • Val pairs: 2,199
  • Effective batch: 16
  • Max length: 256
  • LR: 0.0001
  • Wall-clock: 1:28:29 (1.47 h)
  • Finished: 2026-10-16T15:39:21

Packaged files

configuration_indictrans.py, modeling_indictrans.py, tokenization_indictrans.py, dict.SRC.json, dict.TGT.json, model.SRC, model.TGT, pytorch_model.bin, config.json, tokenizer_config.json, generation_config.json

Eval (200 samples, EN→MR, seed=42)

Model BLEU ChrF METEOR TER COMET BERTF1 Judge/5
base (ai4bharat/indictrans2-en-indic-dist-200M) 35.87 56.52 37.17 203.38 0.6741 81.22 2.47
fine-tuned (this repo) 57.55 66.23 48.98 106.01 0.6917 84.34 2.51
Downloads last month
38
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for anuj1541/indictrans2-en-indic-mr-cidco

Finetuned
(17)
this model

Datasets used to train anuj1541/indictrans2-en-indic-mr-cidco