Downloads · 30 days
8
11% of all-time downloads
mishbahul11/nerds-nits-nllb600m-x-en-ar-finetuned-wmt2026
nerds-nits-nllb600m-x-en-ar-finetuned-wmt2026 is a translation model from mishbahul11. Use it when you need text moved from one language to another. The card lists the license as mit.
This model is a fully fine-tuned version of facebook/nllb-200-distilled-600M, adapted for English→Arabic translation as part of the NERDS-NITS submission to the WMT 2026 Low-Resource Arabic–Asian Language Translation…
Downloads · 30 days
8
11% of all-time downloads
All-time downloads
73
Public
Parameters
615M
2.5 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors2.5 GB · 99%
From the Hugging Face model README
This model is a fully fine-tuned version of facebook/nllb-200-distilled-600M, adapted for English→Arabic translation as part of the NERDS-NITS submission to the WMT 2026 Low-Resource Arabic–Asian Language Translation Shared Task.
This model forms the English→Arabic pivot leg of the NERDS-NITS pivot-based translation framework, used for the Hindi→Arabic and Bengali→Arabic directions, where it is combined with pretrained IndicTrans2 checkpoints handling the Indic→English leg.
AutoTokenizer (inherited from the original NLLB-200-600M checkpoint)facebook/nllb-200-distilled-600MSeq2SeqTrainer, minimizing sequence-to-sequence cross-entropy losseng_Latn, Arabic — arb_ArabThis model was selected as the English→Arabic pivot leg after comparing two fine-tuning strategies: LoRA fine-tuning of NLLB-200-1.3B, and full fine-tuning of NLLB-200-600M. The fully fine-tuned 600M model outperformed the 1.3B LoRA model across BLEU, chrF, and COMET on the held-out development set, and was therefore chosen as the pivot leg for the Indic-to-Arabic direction in this submission.
Limitations:
| Hyperparameter | Value |
|---|---|
| Learning rate | 2e-5 |
| Epochs | 3 |
| Effective batch size | 64 |
| Fine-tuning method | Full fine-tuning |
| Hardware | Single NVIDIA T4 (16GB VRAM) |
| Precision | fp16 / bf16 mixed precision |
| LR scheduler | Cosine |
| Early stopping | Patience = 2 (on validation loss) |
This checkpoint and its Arabic→English counterpart were each fully fine-tuned from the same NLLB-200-600M initialization using identical hyperparameters, so that any performance difference between directions reflects genuine asymmetries in the translation task rather than differences in training configuration.
Fine-tuned on the WMT 2026 Arabic–English parallel corpus, consisting of 19,950 training pairs, with 1,050 dev pairs and 500 devtest pairs.
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
model_name = "mishbahul11/nerds-nits-nllb600m-x-en-ar-finetuned-wmt2026"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
# English -> Arabic
tokenizer.src_lang = "eng_Latn"
inputs = tokenizer("Hello world", return_tensors="pt")
generated_tokens = model.generate(
**inputs,
forced_bos_token_id=tokenizer.convert_tokens_to_ids("arb_Arab"),
num_beams=5,
max_new_tokens=192,
)
print(tokenizer.batch_decode(generated_tokens, skip_special_tokens=True))
# Example: Hindi -> English (IndicTrans2) -> Arabic (this model)
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
hi_en_model_name = "ai4bharat/indictrans2-indic-en-dist-200M"
en_ar_model_name = "mishbahul11/nerds-nits-nllb600m-x-en-ar-finetuned-wmt2026"
# Step 1: Hindi -> English (via IndicTrans2)
# See ai4bharat/indictrans2-indic-en-dist-200M model card for full usage details
english_text = "..." # output of IndicTrans2 Hindi->English step
# Step 2: English -> Arabic (this model)
tokenizer_en_ar = AutoTokenizer.from_pretrained(en_ar_model_name)
model_en_ar = AutoModelForSeq2SeqLM.from_pretrained(en_ar_model_name)
tokenizer_en_ar.src_lang = "eng_Latn"
inputs = tokenizer_en_ar(english_text, return_tensors="pt")
tokens = model_en_ar.generate(
**inputs,
forced_bos_token_id=tokenizer_en_ar.convert_tokens_to_ids("arb_Arab"),
num_beams=5,
)
arabic_text = tokenizer_en_ar.batch_decode(tokens, skip_special_tokens=True)[0]
This model serves as the English→Arabic pivot leg; downstream pivot performance (combined with IndicTrans2's Indic→English leg) for Hindi→Arabic and Bengali→Arabic is reported below.
| Direction | Best System | BLEU | chrF | COMET |
|---|---|---|---|---|
| HI→AR | NLLB + IndicTrans (pivot) | 16.56 | 46.3 | 0.8504 |
| BN→AR | NLLB + IndicTrans (pivot) | 15.73 | 45.86 | 0.8498 |
| Direction | Best System | BLEU | chrF2 | TER | COMET | mBERT |
|---|---|---|---|---|---|---|
| HI→AR | NLLB + IndicTrans (pivot) | 15.5918 | 49.0983 | 75.1866 | 81.9127 | 83.4993 |
| BN→AR | NLLB + IndicTrans (pivot) | 15.0870 | 48.5534 | 75.7664 | 81.6119 | 83.3574 |
Note: challenge test set COMET values are reported by the organizers on a 0–100 scale, while devtest COMET values use the standard 0–1 scale.
NERDS-NITS at WMT 2026: Pivot-Based and Direct Neural Machine Translation for Arabic–Asian Low-Resource Language Pairs
We thank the WMT 2026 shared task organizers for providing the datasets and evaluation infrastructure used in this work.