README.md

metadata

language:
  - en
  - ar
tags:
  - mlx
  - cybersecurity
  - security
  - qwen3
  - saif
  - circuit-aware
  - hayula
library_name: mlx
pipeline_tag: text-generation
license: apache-2.0

SAIF v2 Pro — Circuit-Aware Security Specialist

SAIF (سيف) v2 Pro is a unified security analysis model based on Qwen3-8B, trained with Circuit-Aware LoRA and IACR (Integrated Adapter Circuit Refinement).

Architecture

• Base Model: Qwen3-8B (36 layers)
• Circuit-Aware LoRA: Trained on 387 security analysis samples (rank=8, scale=20)
• IACR Fine-tuning: Integrated Adapter Circuit Refinement (loss 0.033)
• 13 Security Probes: Avg accuracy 98.7%
• Fused Format: MLX native (single 15GB model)

Capabilities

#	Capability	Supported
1	SQL Injection Analysis	✅
2	SSRF Detection	✅
3	XSS Analysis	✅
4	RCE Exploitation	✅
5	IDOR Detection	✅
6	CSRF Analysis	✅
7	LFI Detection	✅
8	Open Redirect	✅
9	CORS Misconfig	✅
10	Auth Bypass	✅
11	Business Logic	✅
12	Race Conditions	✅
13	LLM Prompt Inject	✅

Usage (MLX)

from mlx_lm import load, generate

model, tokenizer = load("BinSaqban/SAIF-v2-Pro")
prompt = "Analyze this SSRF vulnerability: ..."
messages = [{"role": "user", "content": prompt}]
prompt_text = tokenizer.apply_chat_template(messages, tokenize=False)
response = generate(model, tokenizer, prompt=prompt_text, max_tokens=512)
print(response)

Server Mode

mlx_lm.server --model BinSaqban/SAIF-v2-Pro --port 8452

Performance

• Latency: ~7.4s for 200 tokens (27 tok/s on M2 Ultra)
• Accuracy: 98.7% on 13 security probes
• Memory: ~16GB (M2 Ultra 192GB)

Training

Trained on Hayula Labs infrastructure using MLX on M2 Ultra (192GB).

• Circuit-Aware: 500 iterations, train loss 0.927
• IACR: Additional refinement, final loss 0.033

License

Apache 2.0