pushkarsharma/arabic-translation-challenge-adapters

Arabic-Centric Machine Translation — Fine-Tuned Checkpoints

LoRA adapters for Arabic-centric machine translation across six directions (ar↔en, ar↔hi, ar↔ur), covering the models evaluated in the accompanying Arabic-Centric Machine Translation Framework.

Code: https://github.com/p-sharma-7/Arabic-Translation-Challenge

Contents

Each directory is <family>/<direction>/ and holds a PEFT adapter plus the tokenizer files saved alongside it.

Family	Base model	Directions
acegpt/	FreedomIntelligence/AceGPT-v2-8B-Chat	6
fanar/	QCRI/Fanar-1-9B-Instruct	6
gemmax2/	ModelSpace/GemmaX2-28-9B-v0.1	6
hala9b/	hammh0a/Hala-9B	6
translategemma/	google/translategemma-4b-it	6
madlad/	google/madlad400-10b-mt	5 (no ar-ur)
nllb/	facebook/nllb-200-3.3B (full fine-tune)	6 — weights truncated, see below

Directions are ar-en, en-ar, ar-hi, hi-ar, ar-ur, ur-ar.

LoRA configuration

Identical across all families: r=16, alpha=32, dropout=0.05, task_type varying with the base architecture. Target modules follow each architecture:

• Decoder-only (acegpt, fanar, gemmax2, hala9b): q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• T5-style (madlad): q, k, v, o, wi_0, wi_1, wo
• translategemma: the decoder-only set plus fc1, fc2, out_proj

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
from huggingface_hub import snapshot_download

path = snapshot_download(
    "pushkarsharma/arabic-translation-challenge-adapters",
    allow_patterns="fanar/ar-en/*",
)

base = AutoModelForCausalLM.from_pretrained("QCRI/Fanar-1-9B-Instruct", device_map="auto")
model = PeftModel.from_pretrained(base, f"{path}/fanar/ar-en")
tok = AutoTokenizer.from_pretrained(f"{path}/fanar/ar-en")

To reconstruct the layout the code expects, fetch the whole set into finetuned_checkpoints/ at the repo root:

from huggingface_hub import snapshot_download
snapshot_download(
    "pushkarsharma/arabic-translation-challenge-adapters",
    local_dir="finetuned_checkpoints",
)

Several base models are gated — request access on the Hub and authenticate (hf auth login, or set HF_TOKEN) before loading.

Known issue: NLLB checkpoints

The nllb/ directories are included here for completeness, but their weights are truncated and cannot be loaded. Each model-0000*-of-00002.safetensors shard is exactly 5 MiB, while the safetensors header declares 687 tensors requiring ~6.23 GiB per direction — shard 1 alone is missing ~4.99 GB. transformers will fail on them.

What is usable in nllb/: config.json, generation_config.json, model.safetensors.index.json, the tokenizer files, and training_args.bin, which together record the NLLB run's configuration. To obtain working NLLB weights, re-run that fine-tuning pipeline. The six LoRA families above are unaffected.