bachuntr-gemma4-26b-q4km

Fine-tuned Gemma 4 26B-A4B-it (MoE, 4B active params) for automated web security scanning and vulnerability discovery.

Model Details

• Base: google/gemma-4-26b-a4b-it (26B total, 4B active MoE)
• Fine-tuning: LoRA rank 32, alpha 32 on attention + MoE expert layers
• Trainable params: 37M / 25.8B (0.14%)
• Training data: 176 distilled security scanning sessions (discoverer, exploiter, verifier, reporter agents)
• Epochs: 2 (44 steps, batch=8, lr=2e-4 cosine)
• Final loss: 0.46 avg, 0.33 last step
• Max seq length: 8,192 tokens
• Quantization: q4_k_m (5.32 BPW) via llama.cpp

Files

File	Size	Description
bachuntr-gemma4-26b-q4km.gguf	16 GB	q4_k_m quantized GGUF for Ollama/llama.cpp
lora_adapter/	142 MB	Raw LoRA adapter (PEFT) for further fine-tuning

Usage with Ollama

# Download the GGUF file, then create Modelfile:
cat > Modelfile << 'MF'
FROM ./bachuntr-gemma4-26b-q4km.gguf
PARAMETER temperature 0.3
PARAMETER top_p 0.9
PARAMETER num_ctx 8192
SYSTEM "You are bachuntr, an expert web application security scanner. You analyze targets for vulnerabilities including authentication bypass, IDOR, SSRF, XSS, SQLi, business logic flaws, and more. You use browser automation tools to discover and verify security issues."
MF

ollama create bachuntr -f Modelfile
ollama run bachuntr

Training Loss Curve

Step	Loss	Epoch
1	1.073	0.05
5	0.834	0.23
10	0.344	0.50
15	0.484	0.68
22	0.283	1.00
30	0.301	1.36
40	0.331	1.82
44	0.330	2.00

Hardware

• Trained on NVIDIA A100-SXM4-80GB (RunPod)
• Training time: ~56 minutes
• Inference: runs on RTX 3090 24GB with Ollama (q4_k_m)

Part of

bachuntr - AI-powered web application security scanner with multi-agent architecture.