COIL-D/translate-nllb-marathi-hindi

NLLB-200-distilled-600M fine-tuned: Marathi → Hindi

This model is a fine-tuned version of facebook/nllb-200-distilled-600M for Marathi (mar_Deva) to Hindi (hin_Deva) machine translation.

Model description

• Base model: facebook/nllb-200-distilled-600M (distilled NLLB-200, 600M parameters)
• Architecture: M2M100ForConditionalGeneration (encoder-decoder Transformer)
• Source language: Marathi — FLORES-200 code mar_Deva
• Target language: Hindi — FLORES-200 code hin_Deva
• Tokenizer: original NLLB-200 SentencePiece tokenizer (sentencepiece.bpe.model), unchanged from the base model

Intended uses

Single-direction machine translation from Marathi to Hindi. For general-purpose multi-language translation, use the original facebook/nllb-200-distilled-600M instead — this checkpoint is specialized for the mar_Deva → hin_Deva direction only.

How to use

import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_id = "COIL-D/translate-nllb-marathi-hindi"
tokenizer = AutoTokenizer.from_pretrained(model_id, src_lang="mar_Deva")
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

text = "your Marathi input text here"
inputs = tokenizer(text, return_tensors="pt")

translated_tokens = model.generate(
    **inputs,
    forced_bos_token_id=tokenizer.convert_tokens_to_ids("hin_Deva"),
    max_length=200,
)
print(tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0])

Training procedure

Fine-tuned from facebook/nllb-200-distilled-600M with the Hugging Face Trainer.

Training hyperparameters

• learning_rate: 2e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: AdamW (torch fused)
• lr_scheduler_type: linear
• num_epochs: 3
• mixed_precision_training: fp16

Framework versions

• Transformers 4.38.2
• PyTorch

Limitations and bias

This model is trained for a single translation direction (mar_Deva → hin_Deva) and has not been evaluated on out-of-domain text. Translation quality for low-resource languages such as Hindi can vary significantly by domain and sentence length. As with the base NLLB-200 model, it may reflect biases present in its training data.

License

This model is derived from facebook/nllb-200-distilled-600M, which is released under the CC-BY-NC-4.0 license. This fine-tuned checkpoint is distributed under the same license and is intended for non-commercial use.

Citation

If you use this model, please cite:

@misc{singh2026coildindiccentricparallelcorpus,
      title={COILD: An Indic-Centric Parallel Corpus and Benchmark for Machine Translation Across Indian Languages},
      author={Kshetrimayum Boynao Singh and Nitin Kumar Mishra and Palash Pratim Dutta and Atai Waris Khan and Aparna Kaushik and Avinash Kumar and Deeksha and Deepak Kumar and Saroj Kumar Jha and Saloka Sengupta and Anansa Roy and Umalatha Kannoth and Saifulla Samar and Meena Sharma and Manpreet Kaur and Jyoti Sharma and Ashwini Vaidya and Muralikrishna SN and Md Shad Akhtar and Poonam Bansal and Amita Dev and Sanasam Ranbir Singh and Samit Bhattacharya and Tanmoy Chakraborty and Asif Ekbal},
      year={2026},
      eprint={2609.28826},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.28826},
}