Downloads · 30 days
0
mishbahul11/nerds-nits-nllb-ar-id-qlora-finetuned-wmt2026
nerds-nits-nllb-ar-id-qlora-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 cc-by-nc-4.0.
This model is a fine-tuned version of facebook/nllb-200-1.3B, adapted for Arabic↔Indonesian translation as part of the NERDS-NITS submission to the WMT 2026 Low-Resource Arabic–Asian Language Translation Shared Task.
Downloads · 30 days
0
Access
Public
Updated Sep 7, 2026
Repo size
410 MB
Likes
0
Public
Click a slice to open those files.
.safetensors378 MB · 92%
From the Hugging Face model README
This model is a fine-tuned version of facebook/nllb-200-1.3B, adapted for Arabic↔Indonesian translation as part of the NERDS-NITS submission to the WMT 2026 Low-Resource Arabic–Asian Language Translation Shared Task.
r = 64, α = 128 (α = 2r), targeting query, key, value, output, and feed-forward projection layers; ~4.2% of total parameters trained, remainder frozenAutoTokenizer (inherited from the original NLLB-200-1.3B checkpoint)facebook/nllb-200-1.3BSeq2SeqTrainer with Adafactor optimizer, minimizing sequence-to-sequence cross-entropy lossarb_Arab, Indonesian — ind_LatnThis model is intended for research use in low-resource Arabic↔Indonesian machine translation, particularly in the context of the WMT 2026 shared task. It was selected as the direct fine-tuning approach for this pair since IndicTrans2 (used for the pivot-based Arabic–Indic systems in the same submission) does not cover Indonesian.
Limitations:
| Hyperparameter | Value |
|---|---|
| Learning rate | 2e-4 |
| Warmup ratio | 0.05 |
| Epochs | 6 |
| Effective batch size | 32 (batch size 4, gradient accumulation 8) |
| Max sequence length | 192 tokens |
| Optimizer | Adafactor |
| LoRA rank (r) | 64 |
| LoRA alpha (α) | 128 |
| Quantization | NF4 (QLoRA) |
| Hardware | Single NVIDIA T4 (16GB VRAM) |
| Precision | fp16 / bf16 mixed precision |
| Inference decoding | Beam search, beam width 5 |
Fine-tuned on the WMT 2026 Arabic–Indonesian parallel corpus.
| Direction | BLEU | chrF | COMET |
|---|---|---|---|
| AR→ID | 31.97 | 61.25 | 0.8582 |
| Direction | BLEU | chrF2 | TER | COMET | mBERT |
|---|---|---|---|---|---|
| AR→ID | 30.5145 | 59.8177 | 59.1446 | 84.1057 | 86.0613 |
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.
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from peft import PeftModel
base_model_name = "facebook/nllb-200-1.3B"
adapter_name = "mishbahul11/nerds-nits-nllb-ar-id-qlora-finetuned-wmt2026"
tokenizer = AutoTokenizer.from_pretrained(adapter_name)
base_model = AutoModelForSeq2SeqLM.from_pretrained(
base_model_name,
load_in_4bit=True,
device_map="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_name)
# Arabic -> Indonesian
tokenizer.src_lang = "arb_Arab"
inputs = tokenizer("مرحبا بالعالم", return_tensors="pt").to(model.device)
generated_tokens = model.generate(
**inputs,
forced_bos_token_id=tokenizer.convert_tokens_to_ids("ind_Latn"),
num_beams=5,
max_new_tokens=192,
)
print(tokenizer.batch_decode(generated_tokens, skip_special_tokens=True))
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.