🔧 Security Builder Model (14B)

Fine-tuned Qwen2.5-Coder-14B-Instruct khusus untuk generasi patch keamanan & penulisan kode aman. Melengkapi Auditor model dengan mengubah laporan kerentanan menjadi kode perbaikan yang production-ready.

🚀 Quick Load

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "lablab-ai-amd-developer-hackathon/security-builder-14b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")

###  💬 Example Usage (JSON Mode)
messages = [
    {"role": "user", "content": "Fix the buffer overflow and return JSON with keys: fixed_code, explanation, cwe_mitigated."}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=512, temperature=0.1)

import json
print(json.loads(tokenizer.decode(output[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)))

🛠️ Technical Specifications

Parameter	Value
Base Model	Qwen2.5-Coder-14B-Instruct
Fine-tuning	LoRA (r=64, alpha=128, dropout=0.05)
Training Data	Custom secure coding & patch dataset
Epochs	3
Precision	float16 (ROCm-optimized)
Format	Safetensors (6 shards, ~28GB)
VRAM Required	~38-42 GB

🖥️ ROCm & Hardware Optimization

Dioptimalkan untuk AMD Instinct MI300X / ROCm 7.0. Disarankan set env var berikut sebelum inference: export HSA_OVERRIDE_GFX_VERSION=11.0.0 export PYTORCH_HIP_ALLOC_CONF=expandable_segments:False