Downloads · 30 days
657
42% of all-time downloads
mmp2055/Qwen3.5-35B-A3B-uncensored-bughunter-v8
Qwen3.5-35B-A3B-uncensored-bughunter-v8 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.
First MoE iteration of the bughunter series. LoRA fine-tune of llmfan46/Qwen3.5-35B-A3B-uncensored-heretic-v2-Native-MTP-Preserved, specialised in offensive and defensive security work.
Downloads · 30 days
657
42% of all-time downloads
All-time downloads
1.6K
Public
Repo size
22.1 GB
Likes
2
Public
Click a slice to open those files.
.gguf22.1 GB · 100%
From the Hugging Face model README
First MoE iteration of the bughunter series. LoRA fine-tune of llmfan46/Qwen3.5-35B-A3B-uncensored-heretic-v2-Native-MTP-Preserved, specialised in offensive and defensive security work.
v8 vs v7: same dataset and same hyperparameters as bughunter-v7 (27B dense), applied on a 35B MoE base with 3B active params per token. Result: inference speed ~4× better on Apple Silicon (30 tok/s vs 7 tok/s of v7) with comparable quality in 90% of cases.
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:
Qwen3.5-35B-A3B-uncensored-bughunter-v8.Q4_K_M.gguf — the LLM, ~20 GB (Q4_K_M).mmproj-BF16.gguf — the vision projector, ~874 MB (BF16). Download this only if you want image understanding. Place it in the same folder as the LLM (LM Studio auto-detects it because it starts with mmproj-).Vision is inherited from the base and untouched by the fine-tune (the fine-tune corpus is 100% text; the vision tower and projector were not trained). The mmproj-BF16.gguf published here is the one extracted from the base by convert_hf_to_gguf.py.
Runs directly on:
qwen35moe support).qwen35moe architecture.The convert_hf_to_gguf.py converter of llama.cpp does not export the MTP drafter tensors for the qwen35moe arch at the time of this publication (known upstream bug). The GGUF metadata has already been patched with nextn_predict_layers=0 and block_count=40 to reflect that absence, so plain llama.cpp loads without problems.
However, LM Studio detects the qwen35moe arch and automatically enables Speculative Decoding = MTP, ignoring the metadata. This causes a load error (failed to create MTP context). Fix: when loading the model in LM Studio, go to Load → Speculative Decoding → Off. Then Load Model. It should load correctly.
When LM Studio fixes the auto-detect for qwen35moe, this workaround will no longer be needed.
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.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. Same dataset as v6 and v7 — extracted from 1,690 source documents grouped into 9 sources with weighted sampling to prioritise specialised offensive content.
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:
Same configuration as v6/v7 (proven not to trigger 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, shared_expert.{gate,up,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 |
| Trainable parameters | 4,177,920 (0.01% of the 35B total) |
LoRA strategy on MoE (Option B — attention + shared_expert):
q/k/v/o_proj (attention, 10 full-attention layers out of 40) + shared_expert.{gate,up,down}_proj (backbone dense MLP always active, all 40 layers).mlp.gate.weight (expert router), mlp.experts.* (256 individual experts), mtp.* (base drafter).Framework: Unsloth 2026.8.22 on Colab A100 80GB (~10h 20min).
Healthy loss curve, well-behaved cosine scheduler:
| Step | train_loss | eval_loss |
|---|---|---|
| 100 | 2.295 | 2.339 |
| 500 | 1.267 | 1.469 |
| 900 | 0.961 | 1.330 |
| 1300 | 1.169 | 1.291 |
| 1500 | 0.958 | 1.283 |
| final (1865) | 0.947 | 1.280 |
Final average training loss: 1.2522.
The eval_loss decreased monotonically throughout training with a clear plateau in the last third — sign of healthy generalisation, not overfit. Absolute losses are NOT directly comparable to v6/v7 (different bases, different architectures — dense vs MoE).
peft.merge_and_unload() on a local Mac (32 GB RAM).convert_hf_to_gguf.py + llama-quantize (llama.cpp master).mmproj) extracted separately in BF16 (874 MB).block_count 41 → 40, nextn_predict_layers 1 → 0) to reflect that the drafter's MTP tensors were not exported..Q4_K_M.gguf and mmproj-BF16.gguf) into the same folder.~/.lmstudio/models/<publisher>/<repo>/.Load tab → Speculative Decoding → Off. If you leave it at the default MTP, LM Studio will try to activate an MTP drafter that this GGUF does not contain and will fail with failed to create MTP context.qwen35moe doesn't tolerate it well).llama-server \
-m Qwen3.5-35B-A3B-uncensored-bughunter-v8.Q4_K_M.gguf \
--mmproj mmproj-BF16.gguf \
-c 32768 \
-fa on \
-ngl 999 \
--port 8080
OpenAI-compatible endpoint at http://127.0.0.1:8080/v1/chat/completions.
Bughunter Harness is a multi-agent orchestrated pipeline that drives this model through a full 9-stage bug bounty workflow — from reconnaissance to consolidated Markdown report — with rate limits, scope gates and secret redaction enforced by default. This model (v8) is the recommended and pinned backend in the harness config.
Pipeline (9 agents run in sequence, order matters):
recon — subdomain + host + port enumeration.fingerprint — tech-stack detection.content_discovery — hidden paths + historical URLs (gau/wayback/katana).login_probe — LAB ONLY (localhost / private IP / DVWA / Juice Shop / Mutillidae / bWAPP / WebGoat) — 6 stock default-cred pairs, session cookie harvest.web_vuln — nuclei + nikto + sqlmap + dalfox on every parameterized endpoint found in step 3, cookie from step 4 auto-injected.wordpress — wpscan when WordPress is detected.api_fuzzer — API surface + BOLA/BFLA hints.auth — auth-bypass + SSO/OAuth misconfig.report — deterministic Markdown consolidator → sessions/<run-id>/REPORT.md.Design goals — safe-by-default, model-agnostic:
rm, drop, wipe, …).Backend support: LM Studio, Ollama, Llama.cpp (all local) — plus OpenAI, Anthropic, NVIDIA NIM, Google Gemini (cloud) via --servertype. Local is recommended for privacy + zero cost.
Setup with LM Studio (fastest path):
mmp2055/Qwen3.5-35B-A3B-uncensored-bughunter-v8 (~19 GB Q4_K_M + optional mmproj).qwen35moe today).run_shell, http_get, etc.).Install the harness:
git clone https://github.com/marcodemm/bughunter-harness.git
cd bughunter-harness
python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
# Copy templates
cp config.example.yaml config.yaml
cp scope.example.txt scope.txt
# Edit config.yaml (backend, SMTP/Telegram optional, custom_headers, …)
# Edit scope.txt with hostnames you are authorized to test
The default llm.model in config.example.yaml is already pinned to this v8 model — no config change needed if you follow the LM Studio setup above.
Quick start:
source .venv/bin/activate
python harness.py
At the prompt:
Objective (one-line goal for the agent). Type /quit or /bye to exit.
Inline flags (sticky): --email ADDR --scope PAT (repeat) -o PATH
> http://localhost:3000/
The orchestrator runs the 9 agents in sequence and produces sessions/<run-id>/REPORT.md at the end.
Optional tooling (harness auto-detects and uses whatever is installed):
# macOS Homebrew — most useful subset
brew install nuclei ffuf httpx subfinder katana nmap nikto \
feroxbuster wpscan dnsx naabu jq sqlmap
brew install gau waybackurls # historical URL discovery
go install github.com/hahwul/dalfox/v2@latest # active XSS scanner
nuclei -update-templates # first run
Plus SecLists at /opt/homebrew/share/seclists for content_discovery and api_fuzzer wordlists.
Mandatory operational rules for any offensive use:
scope.txt allowlist. No exceptions.-rl 2 -c 1, dalfox --delay 1000 -w 1, ffuf --rate 2, nikto -Pause 1.<project>-<vector>-NN.whoami for RCE, 4-char redacted for secrets, 2 own accounts for IDOR, ≤20 requests for races.Troubleshooting:
| Symptom | Likely cause | Fix |
|---|---|---|
| LM Studio fails to load, MTP error | Speculative Decoding auto-enabled | Warning at top of this README — disable Speculative Decoding in Model Settings, reload |
| Harness agents don't execute any tool | Tool Use disabled in LM Studio | Model Settings → Tool Use ON (mandatory for the agent pipeline) |
curl 127.0.0.1:1234/v1/models returns Connection refused | LM Studio Server not started | LM Studio → Developer → Local Server → Start Server |
| Agents time out on the first tool call | Local model too slow for pipeline latency | Reduce context to 16 k, max GPU Offload, close apps competing for RAM |
| Model replies in English when it should be in Spanish | System prompt not applied | bughunter-v8 defaults to Spanish; verify the chat template is being applied (LM Studio → Chat tab → try a turn) |
| Replies hallucinate non-existent tools/flags | Insufficient context for the task | Raise Context Length in LM Studio or reduce the session history |
| Very slow (>30 s per reply) | Large context + Q4_K_M + MoE = memory-bandwidth bottleneck | Reduce context to 16 k, max GPU Offload, close apps competing for RAM |
Agent proposes destructive commands (rm, drop, wipe) | Not possible with the harness — destructive verbs are blocked at the tool-call layer | If seen, open an issue on the harness repo — should not happen |
References:
config.example.yaml.content_discovery and api_fuzzer.You can use the model without a system prompt (it behaves correctly by default). For bug bounty tasks specifically:
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, 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 (~17 MB, 4.2M params)
├── chat_template.jinja
├── tokenizer_config.json
├── tokenizer.json
└── README.md
Merge into the base and use as a standalone model:
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel
from huggingface_hub import snapshot_download
BASE = "llmfan46/Qwen3.5-35B-A3B-uncensored-heretic-v2-Native-MTP-Preserved"
adapter_dir = snapshot_download(
"mmp2055/Qwen3.5-35B-A3B-uncensored-bughunter-v8",
allow_patterns="adapter/*",
)
base = AutoModelForImageTextToText.from_pretrained(BASE, dtype=torch.bfloat16, device_map="cpu", trust_remote_code=True)
model = PeftModel.from_pretrained(base, f"{adapter_dir}/adapter")
model = model.merge_and_unload()
processor = AutoProcessor.from_pretrained(BASE, trust_remote_code=True)
qwen35moe arch. Without the drafter, there is no speculative-decoding speedup (the model runs at ~30 tok/s on Apple Silicon M5 instead of the ~50-60 that MTP would enable).This model is distributed for:
It is not distributed for:
The user assumes all legal responsibility for the use of the model.
qwen35moe, 256 experts with 8 active per token + dense shared expert, hybrid attention Gated DeltaNet + full attention, context up to 262,144 tokens, native Multi-Token Prediction).qwen3_5_moe.@misc{qwen35moe_bughunter_v8,
title = {Qwen3.5-35B-A3B-uncensored-bughunter-v8: MoE fine-tune specialised in bug bounty and offensive/defensive security},
year = {2026},
base = {llmfan46/Qwen3.5-35B-A3B-uncensored-heretic-v2-Native-MTP-Preserved},
method = {LoRA fine-tune (rank=8, lr=1e-5, 1 epoch, QLoRA 4-bit, Unsloth) on attention + shared_expert layers},
data = {14914 ShareGPT examples distilled from public bug bounty corpus}
}