Downloads · 30 days
1K
55% of all-time downloads
mmp2055/qwen3.5-9b-claude-4.6-os-av-h-uncensored-think-d_au-imat-bughunter-v6
qwen3.5-9b-claude-4.6-os-av-h-uncensored-think-d_au-imat-bughunter-v6 is a image-text-to-text model from mmp2055. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for gguf. The card lists the license as apache-2.0.
Fine-tune of Qwen3.5-9B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING by DavidAU, specialised in offensive and defensive security work. Trained on a distilled corpus of 15,611 knowledge units covering bug bo…
Downloads · 30 days
1K
55% of all-time downloads
All-time downloads
1.8K
Public
Repo size
7.5 GB
Likes
1
Public
Click a slice to open those files.
.gguf7.5 GB · 99%
From the Hugging Face model README
Fine-tune of Qwen3.5-9B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING by DavidAU, specialised in offensive and defensive security work. Trained on a distilled corpus of 15,611 knowledge units covering bug bounty, web pentesting, traffic analysis, DFIR, memory forensics and incident response.
Designed to assist researchers in real engagements on bug bounty platforms (HackerOne, Bugcrowd, Intigriti, YesWeHack, Immunefi) with operational rules baked in: authorized scope only, minimal non-destructive PoCs, callbacks to owned infrastructure. Answers in Spanish; payloads, commands and report templates are in English.
Two files:
qwen3.5-9b-claude-4.6-os-av-h-uncensored-think-d_au-imat-bughunter-v6.Q4_K_M.gguf — the LLM, ~5.3 GB (Q4_K_M with calibrated imatrix).mmproj-F32.gguf — the vision projector, ~1.7 GB. Download this only if you want image understanding. Place it in the same folder as the LLM.Vision is inherited from the DavidAU base and untouched by the fine-tune (the vision tower and projector were not trained). The mmproj-F32.gguf published here is the same one DavidAU ships with the base, so it works identically. Do note that our fine-tune corpus is 100% text, so the LLM is not specifically tuned to reason over images — it can describe them and answer basic visual questions, but its added value is on the text side.
Runs directly on:
Covers the full lifecycle of a bug bounty engagement:
{{7*7}} on Jinja/Twig/Freemarker/Velocity), RCE via deserialization, XXE in SAML/SOAP/SVG/OOXML.It also covers specialised areas: WiFi security (WPA/WPA2/WPA3, evil twin, deauth), Sub-GHz and SDR analysis, binary reverse engineering, exploit development, MITRE ATT&CK techniques.
When a system prompt is provided, the model incorporates ethical research rules:
whoami/id for RCE; SELECT version() for SQLi; two owner accounts for IDOR.14,914 examples in ShareGPT format (train), 466 valid, 309 test. Extracted from 1,690 source documents grouped into 9 sources with weighted sampling to prioritise specialised offensive content:
| Source | Weight | % in train |
|---|---|---|
| Patterns distilled from real HackerOne reports | ×4 | 7.9% |
| Reproducible PortSwigger Web Security Academy methods | ×3 | 6.6% |
| Factual HackerOne reports (by category) | ×3 | 3.5% |
| Researcher's own tooling | ×3 | 12.9% |
| Lab walkthroughs | ×2 | 5.5% |
| Structured offensive skills | ×2 | 20.8% |
| Mixed skills (offensive + defensive) | ×1.5 | 12.1% |
| CTF write-ups | ×1 | 8.6% |
| Defensive / governance / patch management skills | ×0.35 | 21.6% |
Additional: 72 short conversational examples (greetings, meta-questions, ambiguous requests that get answered by asking back) to prevent format collapse — the model learns to modulate length based on the question rather than dumping methodology for any input.
Three-thirds system prompt setup:
This prevents the model from becoming dependent on the exact wording of a specific system prompt.
Key configuration aligned with what worked in similar prior work (learned after several previous attempts that failed due to format collapse):
| Parameter | Value |
|---|---|
| Learning rate | 1e-5 (low; avoids format memorisation) |
| Epochs | 1 (0.54 effective epochs — learn without overfitting) |
| LoRA rank | 8 |
| LoRA alpha | 16 |
| LoRA dropout | 0.05 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Training quantization | QLoRA 4-bit (bitsandbytes) |
| Effective batch | 8 (per_device 1 × grad_accum 8) |
| Cutoff length | 4096 |
| Scheduler | Cosine with 3% warmup |
| Total steps | 1,865 |
| Precision | bf16 |
Framework: LLaMA-Factory on Colab A100 80GB (~5h 15min).
Healthy loss curve, no overfitting:
| Step | train_loss | eval_loss |
|---|---|---|
| 100 | ~1.95 | 1.893 |
| 400 | ~1.75 | 1.698 |
| 700 | ~1.65 | 1.630 |
| 1000 | ~1.60 | 1.588 |
| 1500 | ~1.57 | 1.561 |
| 1600 | ~1.55 | 1.559 |
| final (1865) | 1.6783 | ~1.55 |
The eval_loss stayed at or below train_loss throughout, indicating real learning without formatting overfit.
llamafactory-cli export).convert_hf_to_gguf.py from llama.cpp.system messages (compatibility with clients that send their own system + an additional one, common in AI-enabled IDEs).The source corpus went through blocking PII auditing in two layers (extraction and final build). Sanitised:
*.oastify.com, *.burpcollaborator.net)Public lab credentials (like PortSwigger's wiener:peter) were kept on purpose, since they are part of the documented technique.
.Q4_K_M.gguf and mmproj-F32.gguf) into the same folder.~/.lmstudio/models/<folder>/.On an Apple Silicon Mac with 32 GB RAM, the model runs the full 262,144-token context with room to spare.
llama-server \
-m qwen3.5-9b-claude-4.6-os-av-h-uncensored-think-d_au-imat-bughunter-v6.Q4_K_M.gguf \
--mmproj mmproj-F32.gguf \
-c 65536 \
-ctk q8_0 -ctv q8_0 \
-fa on \
-ngl 999 \
--port 8080
OpenAI-compatible endpoint at http://127.0.0.1:8080/v1/chat/completions.
You can use the model without a system prompt (it behaves correctly by default). For bug bounty tasks specifically, the recommended prompt is:
You are a professional bug bounty researcher. You work on HackerOne, Bugcrowd, Intigriti, YesWeHack, Immunefi and independent programs.
Inviolable rules:
- In-scope assets only, verified against the program scope before every request.
- Sustained throughput ≤1-2 req/s.
- Minimal non-destructive PoC.
- OOB callbacks always to owned infrastructure.
- Never use credentials or tokens found in a finding to validate access.
- Minimum reportable severity: High or Critical.
Answer in Spanish. Payloads, commands and report templates in English.
/adapterIf you want to build on top of this fine-tune — for example adding another language, composing with a different domain LoRA, or re-merging with different quantization settings — the LoRA adapter used to train this model is published inside this same repo under /adapter:
adapter/
├── adapter_config.json # PEFT config: rank 8, alpha 16, dropout 0.05
├── adapter_model.safetensors # LoRA delta weights (~56 MB, 256 tensors)
├── chat_template.jinja # Patched template (accepts multiple system messages)
├── tokenizer_config.json
├── tokenizer.json
└── processor_config.json
Option A — merge into the base and use as a standalone model:
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
from huggingface_hub import snapshot_download
BASE = "DavidAU/Qwen3.5-9B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING"
adapter_dir = snapshot_download(
"mmp2055/qwen3.5-9b-claude-4.6-os-av-h-uncensored-think-d_au-imat-bughunter-v6",
allow_patterns="adapter/*",
)
base = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(base, f"{adapter_dir}/adapter")
model = model.merge_and_unload()
tokenizer = AutoTokenizer.from_pretrained(BASE, trust_remote_code=True)
Option B — keep as adapter and compose with another LoRA:
model = PeftModel.from_pretrained(base, f"{adapter_dir}/adapter", adapter_name="bughunter")
model.load_adapter("your-org/your-language-lora", adapter_name="lang")
model.set_adapter(["bughunter", "lang"]) # both active
Notes when composing:
chat_template.jinja if you rely on IDE agents that inject their own system messages; otherwise the DavidAU original works fine.temperature=0.1 and top_p=0.9 on the client to compensate.This model is distributed for:
It is not distributed for:
The user assumes all legal responsibility for the use of the model.
@misc{qwen35_bughunter_v6,
title = {qwen3.5-9b-claude-4.6-os-av-h-uncensored-think-d_au-imat-bughunter-v6: Fine-tune specialised in bug bounty and offensive/defensive security},
year = {2026},
base = {DavidAU/Qwen3.5-9B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING},
method = {LoRA fine-tune, rank=8, lr=1e-5, 1 epoch, QLoRA 4-bit},
data = {14914 ShareGPT examples distilled from public bug bounty corpus}
}