Fg-selective-arabic.bin

uvicorn main:app --host 0.0.0.0 --port 8000 --workers 2 Now you have a ready for internal tools, chat‑bots, or research pipelines. 6. Performance Benchmarks & Comparative Evaluation | Metric | Fg-selective-arabic.bin | GPT‑4‑Turbo (Arabic) | LLaMA‑2‑13B‑Arabic | MPT‑7B‑Arabic | |--------|---------------------------|---------------------|-------------------|---------------| | Perplexity (MSA) | 13.7 | 13.9 | 16.4 | 19.1 | | BLEU (Summarization) | 35.2 | 34.8 | 30.7 | 28.3 | | ROUGE‑L (QA) | 48.5 | 48.1 | 44.0 | 41.6 | | Inference Latency (RTX 4090, 1‑token) | 9 ms | 12 ms | 13 ms | 15 ms | | VRAM Footprint (FP16) | 7.8 GB | 9.2 GB | 9.8 GB | 8.6 GB | | Dialectal Accuracy (Egyptian) | 92 % | 90 % | 84 % | 80 % |

model_path = "fg-selective-arabic.bin" tokenizer = AutoTokenizer.from_pretrained("fg-consortium/fg-selective-arabic", trust_remote_code=True) Fg-selective-arabic.bin

@app.post("/generate") async def generate(req: GenerationRequest): text = generate_arabic( req.prompt, max_new_tokens=req.max_new_tokens, temperature=req.temperature, top_p=req.top_p ) return "generated_text": text Run with: uvicorn main:app --host 0

# Load with `torch_dtype` set for mixed‑precision model = AutoModelForCausalLM.from_pretrained( model_path, device_map="auto", torch_dtype=torch.bfloat16, # use bfloat16 on Ampere+ GPUs trust_remote_code=True ) model.eval() def generate_arabic(prompt, max_new_tokens=150, temperature=0.8, top_p=0.95): inputs = tokenizer(prompt, return_tensors="pt").to(model.device) with torch.no_grad(): output = model.generate( **inputs, max_new_tokens=max_new_tokens, temperature=temperature, top_p=top_p, do_sample=True, pad_token_id=tokenizer.eos_token_id ) return tokenizer.decode(output[0], skip_special_tokens=True) This article unpacks everything a practitioner

# Example usage prompt = "اكتب مقالًا قصيرًا عن تأثير الذكاء الاصطناعي على التعليم في العالم العربي" print(generate_arabic(prompt)) from fastapi import FastAPI, Request from pydantic import BaseModel

One of the most noteworthy contributions to the Arabic NLP community in 2025 is the checkpoint—a compact, fine‑tuned binary released by the Focal‑Gating (FG) research consortium . This article unpacks everything a practitioner, researcher, or hobbyist needs to know about this file: its origins, internals, practical deployment, performance, and the broader implications for Arabic AI. 2. What Is “Fg‑selective‑arabic.bin”? | Attribute | Description | |-----------|-------------| | File type | Serialized PyTorch checkpoint ( .bin ) | | Model family | Focal‑Gating (FG) Transformer, 1.3 B parameters | | Training regime | Selective fine‑tuning on a curated Arabic corpus (≈ 200 B tokens) | | Primary purpose | High‑quality Arabic text generation, summarization, and instruction following | | Target hardware | GPU‑accelerated inference (≥ 8 GB VRAM) and optional CPU‑only inference via GGUF conversion | | License | Apache 2.0 with a “non‑commercial‑use” addendum (see Section 10) | | Release date | 3 March 2025 (v1.0) | | Version | v1.0‑selective‑2025‑03 (semantic versioning) |