Thaka-AI/aisa-arabicfc-lora

Thaka AISA-ArabicFC adapter

LoRA adapter for the AISA-ArabicFC shared task (ArabicNLP 2026). It fine-tunes google/gemma-4-12B-it for Arabic function calling: given a user request and a candidate tool catalog, emit one tool call with arguments, or abstain.

• Base: google/gemma-4-12B-it
• Type: LoRA, rank 64, bfloat16
• Training data: the provided AISA-ArabicFC training split only

Use

Serve with vLLM:

vllm serve google/gemma-4-12B-it \
    --enable-lora --max-lora-rank 64 \
    --lora-modules primary=./ \
    --served-model-name primary --dtype bfloat16 --max-model-len 6144