ISLAM-PO/MasryGPT-Flash-Adapter

MasryGPT Flash — LoRA Adapter

Try the adapter: Follow TESTING.md and load it on top of the exact base model.

Evaluation status: See EVALUATION.md for the adapter smoke-test plan and reproducibility limits.

Overview

This repository contains a LoRA/PEFT adapter for Arabic text generation. It is not a standalone checkpoint: the adapter is applied on top of the verified base model:

aubmindlab/aragpt2-base

Repository contents

• adapter_config.json — PEFT configuration; base model is now correctly set.
• adapter_model.safetensors — LoRA adapter weights.
• config.json — architecture metadata copied from the base model for reference.
• Tokenizer and chat-template files.
• LICENSE — the AraGPT2 license notice.

Load the adapter

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "aubmindlab/aragpt2-base"
adapter_id = "ISLAM-PO/MasryGPT-Flash-Adapter"

base = AutoModelForCausalLM.from_pretrained(base_id)
model = PeftModel.from_pretrained(base, adapter_id)
tokenizer = AutoTokenizer.from_pretrained(adapter_id)

Important

Do not load this repository as a standalone AutoModelForCausalLM; the base-model weights are intentionally stored in the base-model repository. The adapter must be compatible with the exact tokenizer and architecture of aubmindlab/aragpt2-base.

License

The adapter is distributed under the AraGPT2 license reproduced in LICENSE. Review the base-model and training-data terms before redistribution or commercial use.

---

Metadata corrected to match the actual repository contents.