This model was trained on a curated parallel corpus of 49k Hindi-Kangri sentence pairs, with additional vocabulary and tokenizer extensions for kang_Deva support.

Model Details

Model Architecture: Transformer (Encoder-Decoder)
Base: facebook/nllb-200-distilled-600M
Languages:
- Source: hin_Deva (Hindi)
- Target: kang_Deva (Kangri in Devanagari script)
Tokenizer: SentencePiece with extended vocabulary for kang_Deva
Direction Supported: Hindi → Kangri only (unidirectional)

How to Use

python

import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline

model_name = "cloghost/nllb-200-distilled-600M-hin-kang-v2"
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

device = 0 if torch.cuda.is_available() else -1 
translator = pipeline(
    "translation",
    model=model,
    tokenizer=tokenizer,
    src_lang="hin_Deva",
    tgt_lang="kang_Deva",
    device=device
)

text = """मगर हिमाचली भाषा तो पहले से बोली जा रही है।
लोग सदियों से ही इसके संग जी रहे हैं।
पहाड़ी भाषा का इतिहास हिन्दी साहित्य के आदिकाल ,‌जिसे सिद्ध चारण काल के नाम से भी जानते हैं
"""

translation = translator(text)

Benchmark Scores

Evaluated on a clean 5k sample held-out test set:

Metric	Value
BLEU	26.03
BLEU-4	14.11
ROUGE-1	4.73%
ROUGE-L	4.76%
METEOR	43.63%
BERTScore-F1	93.39%
BERT Precision	93.42%
BERT Recall	93.37%
ChrF	53.93
TER (↓ is better)	56.96
Empty Predictions	0

rayvex/nllb-200-distilled-600M-hin-kang-v2

NLLB-200 Distilled 600M — Hindi → Kangri (v2)

Model Details

How to Use

Benchmark Scores

rayvex/
nllb-200-distilled-600M-hin-kang-v2