Correct model card for adapter-only repository

Document the missing base model and prevent the adapter from being mistaken for a standalone model.

README.md

@@ -1,91 +1,55 @@
-# AraGPT2-Fresh-Forgotten Model
-
-## Overview
-
-**aragpt2-fresh-forgotten** is an Arabic language model built on the **GPT-2 Small** architecture by OpenAI, trained for Arabic text generation.
-
-- **Type:** `GPT2LMHeadModel` (causal language modeling)
-- **Language:** Arabic (with support for Latin characters, digits, and punctuation)
-- **Size:** ~135M parameters (~515 MB in `safetensors` FP32 format)
-- **Vocabulary:** 64,000 tokens (ByteLevel-BPE)
-- **Context length:** 1024 tokens
-
-## Folder Contents
-
-| File | Approx. Size | Description |
-|---|---|---|
-| `config.json` | 881 bytes | Architecture and hyperparameter settings |
-| `generation_config.json` | 204 bytes | Default generation settings |
-| `model.safetensors` | ~515 MB | Model weights |
-| `tokenizer.json` | ~4.6 MB | Tokenizer (BPE) |
-| `vocab.json` | ~1.9 MB | Vocabulary (64k) |
-| `merges.txt` | ~1.5 MB | Merge rules (63,740 merges) |
-| `README.md` | — | This document |
-| `ARCHITECTURE.md` | — | Detailed architecture description |
-| `USAGE.md` | — | Usage guide with code examples |
-| `TOKENIZER.md` | — | Tokenizer and vocabulary guide |
-
-## Quick Specs (from `config.json`)
-
-```json
-{
-  "model_type": "gpt2",
-  "architectures": ["GPT2LMHeadModel"],
-  "n_layer": 12,
-  "n_head": 12,
-  "n_embd": 768,
-  "n_positions": 1024,
-  "n_ctx": 1024,
-  "vocab_size": 64000,
-  "activation_function": "gelu_new",
-  "bos_token_id": 0,
-  "eos_token_id": 0
-}
-```
-
-## Quickstart
-
-```python
-from transformers import GPT2LMHeadModel, GPT2Tokenizer
-
-model_path = r"C:\Users\RDP\Desktop\aragpt2-fresh-forgotten"
-tokenizer = GPT2Tokenizer.from_pretrained(model_path)
-model = GPT2LMHeadModel.from_pretrained(model_path)
-
-prompt = "Artificial intelligence in the Arabic language"
-inputs = tokenizer(prompt, return_tensors="pt")
-outputs = model.generate(
-    **inputs,
-    max_length=100,
-    do_sample=True,
-    top_p=0.95,
-    num_beams=5,
-    repetition_penalty=3.0,
-    no_repeat_ngram_size=3
-)
-print(tokenizer.decode(outputs[0], skip_special_tokens=True))
-```
-
-> For full details see: `USAGE.md`, `ARCHITECTURE.md`, `TOKENIZER.md`
-
-## Intended Uses
-
-- Arabic text generation (stories, articles, sentence completion).
-- Research and experimentation on Arabic language modeling.
-- Fine-tuning for tasks such as dialogue, summarization, and generative classification.
-
-## Limitations and Cautions
-
-1. **Hallucination:** May generate incorrect information — verify facts independently.
-2. **Limited context:** Only 1024 tokens (~700–900 Arabic words approx.).
-3. **Small size:** Only 12 layers — less capable than modern large models.
-4. **Bias:** Reflects biases in the Arabic training data.
-5. **Start/end tokens:** `bos_token_id = eos_token_id = 0` (`<|endoftext|>`) — take care during fine-tuning.
-
-## License and Source
-
-- Architecture: GPT-2 (OpenAI).
-- Check the original training-data license before commercial use.
+---
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- peft
+- lora
+- adapter
+- arabic
+- text-generation
+---
+
+# MasryGPT Flash — LoRA Adapter
+
+## Status
+
+This repository contains a **LoRA/PEFT adapter only**, not a standalone language model.
+
+The repository currently includes:
+
+- `adapter_config.json`
+- `adapter_model.safetensors`
+- Tokenizer support files
+- Chat template
+
+It does **not** include:
+
+- A base-model `config.json`
+- Base-model weights
+- A complete standalone `model.safetensors`
+- A verified base-model repository ID
+
+## Important issue to fix before loading
+
+The current `base_model_name_or_path` points to `ISLAM-PO/masrygpt-flash`, which is this adapter repository itself. That value must be replaced with the exact Hugging Face repository ID of the base model used during training.
+
+Do not report this repository as a standalone model until the base checkpoint and adapter compatibility are verified.
+
+## Required setup
+
+1. Identify the exact base model and upload it or identify its existing Hub ID.
+2. Set `base_model_name_or_path` in `adapter_config.json` to that exact ID.
+3. Load the base model with `transformers`.
+4. Load this adapter with `peft.PeftModel`.
+
+## Intended use
+
+Arabic text-generation experiments after the base model and adapter have been verified.
+
+## License
+
+The license must follow the licenses of both the base model and the training data. Verify them before redistribution or commercial use.
 
 ---
-*Automatically documented from the model files on 2026-09-03.*
+
+*This card was corrected to match the files currently present in the repository.*