harshbheem/kreol-bhojpuri-translator

Kreol Morisien → Bhojpuri Morisien (Romanized) Translator

Model Description

This repository contains a fine-tuned multilingual sequence-to-sequence translation model built using Hugging Face Transformers. It translates Kreol Morisien (Latin script) to Bhojpuri Morisien (romanized, Latin script).

Language Pair

• Source (NLLB tag): mfe_Latn
• Target (NLLB tag): bho_Latn

Fixed Decoding Configuration (IMPORTANT)

All evaluations and demos use the same decoding configuration:

• num_beams = 6
• no_repeat_ngram_size = 3
• length_penalty = 1.0
• max_length = 96
• early_stopping = True

Language forcing:

• tokenizer.src_lang = "mfe_Latn"
• tokenizer.tgt_lang = "bho_Latn"
• forced_bos_token_id = tokenizer.convert_tokens_to_ids("bho_Latn")

Evaluation

Metrics computed using sacrebleu on a fixed test split:

• BLEU: 38.55
• chrF: 58.89

Usage Example

import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

repo_id = "harshbheem/kreol-bhojpuri-translator"

tokenizer = AutoTokenizer.from_pretrained(repo_id, use_fast=False)
model = AutoModelForSeq2SeqLM.from_pretrained(repo_id)

tokenizer.src_lang = "mfe_Latn"
tokenizer.tgt_lang = "bho_Latn"
tgt_id = tokenizer.convert_tokens_to_ids("bho_Latn")

decode_kwargs = dict(
    num_beams=6,
    no_repeat_ngram_size=3,
    length_penalty=1.0,
    max_length=96,
    early_stopping=True,
    forced_bos_token_id=tgt_id,
)

text = "mo pe al lekol"
inputs = tokenizer(text, return_tensors="pt")

with torch.inference_mode():
    output = model.generate(**inputs, **decode_kwargs)

print(tokenizer.decode(output[0], skip_special_tokens=True))

Limitations

• Output quality depends on training data coverage and consistency.
• This is a research prototype for academic demonstration; it may produce grammatical errors.
• CPU inference (e.g., free Spaces) may be slower than GPU.

Citation

Harsh Bheem. (2026). Kreol Morisien → Bhojpuri Morisien Neural Machine Translation (Undergraduate dissertation project).