Table | |
|---|
| Developer | athulkrishnan (independent) |
| Model type | Auto-regressive transformer (decoder-only), instruction-tuned |
| Base model | Qwen/Qwen2.5-Coder-14B-Instruct (~14.7B params, 48 layers) |
| Fine-tune method | QLoRA SFT (4-bit NF4 base, LoRA r=32) via Unsloth + TRL |
| Specialisation | Bug-bounty finding triage/validation · recon attack-surface ranking |
| Language | English |
| Context length | 32,768 native (up to 131K with YaRN); trained at 2,048 |
| Precision / formats | Merged BF16 safetensors · Q4_K_M GGUF in gguf/ |
| License | Apache-2.0 (inherited from the Qwen base) |
| Status | Static, offline fine-tune · v1 (see Versions) |
Intended use
Intended use cases
- Finding triage & validation — decide submit vs. kill, sanity-check severity, reason
about real-world impact, and cut duplicate / informational / out-of-scope noise before a
human writes a report.
- Recon prioritisation — turn a fingerprinted tech stack or attack surface into a ranked
hit-list of vulnerability classes worth testing first, with one-line rationale.
- Methodology assistant — explain bug classes, CWE mappings, and report framing to support
authorized learning and assessment work.
Downstream use
- A local triage/ranking step inside an authorized bug-bounty or pentest workflow
(human-in-the-loop), e.g. pre-filtering scanner output or drafting impact statements.
- A base for further domain fine-tuning or for pairing with retrieval (RAG) over fresh
CVEs / current program scope.
Out-of-scope and prohibited use
- Testing, scanning, or exploiting systems you are not explicitly authorized to assess.
- Autonomous attack execution without human review — BountyHound is a co-pilot, not an agent.
- Generating malware, phishing, or weaponised exploit payloads for unauthorized use.
- Treating outputs as ground truth, or as legal/compliance advice. Always validate.
- Any use that violates applicable law or platform/program rules.
How to get started
Requirements
transformers >= 4.40 (developed on 4.56.2), torch >= 2.3, and accelerate.
The merged model is BF16 (~29 GB); for a single 16 GB GPU use the Q4_K_M GGUF with
llama.cpp / Ollama, or load in 4-bit with bitsandbytes.
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "athulkrishnan/BountyHound-Coder-14B"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
SYSTEM = (
"You are a bug-bounty co-pilot for an authorized security researcher. You assist ONLY "
"with testing that is in-scope and authorized on bug-bounty programs. You are sharp, "
"terse, and impact-first: you kill weak findings, prove real exploitation, and never pad "
"reports with 'could potentially'. Your specialties are finding triage/validation and "
"recon attack-surface ranking."
)
messages = [
{"role": "system", "content": SYSTEM},
{"role": "user", "content":
"Triage: reflected XSS on a marketing page, unauthenticated, no session context. "
"Submit or kill? One line + why."},
]
ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=256, temperature=0.3, top_p=0.9)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))
Ollama / llama.cpp (GGUF)
Download gguf/BountyHound-Coder-14B-Q4_K_M.gguf, then create a Modelfile:
FROM ./BountyHound-Coder-14B-Q4_K_M.gguf
TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{- range .Messages }}<|im_start|>{{ .Role }}
{{ .Content }}<|im_end|>
{{ end }}<|im_start|>assistant
"""
SYSTEM """You are a bug-bounty co-pilot for an authorized security researcher. You assist ONLY with testing that is in-scope and authorized on bug-bounty programs. You are sharp, terse, and impact-first: you kill weak findings, prove real exploitation, and never pad reports with 'could potentially'. Your specialties are finding triage/validation and recon attack-surface ranking."""
PARAMETER temperature 0.3
PARAMETER top_p 0.9
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
ollama create bountyhound -f Modelfile
ollama run bountyhound "Rank the attack surface for a Spring Boot + GraphQL + S3 stack."
Qwen2.5 ChatML (<|im_start|>role … <|im_end|>) with the security system prompt above.
Recommended decoding: temperature 0.3, top_p 0.9, repeat_penalty 1.05.
Training
Training data
A weighted instruction mix biased toward the two target skills (≈6.2K curated conversations):
Table with columns: Source, Purpose| Source | Purpose |
|---|
| HackerOne disclosed reports (public) | finding disposition + severity-triage signal |
| Curated bug-bounty methodology & triage heuristics | submit/kill discipline, validation gates, anti-patterns |
| Recon playbook / attack-surface examples | tech-stack to ranked vulnerability classes |
| Public detection-template patterns | low-false-positive authoring style |
| General-security instruction data (~13%) | rehearsal to limit catastrophic forgetting |
No customer data, private program scope, credentials, or other non-public material is
included in the training set. Only public or self-authored content was used.
Training procedure
QLoRA supervised fine-tuning, loss computed on assistant turns only.
Table with columns: Hyperparameter, Value| Hyperparameter | Value |
|---|
| Quantisation | 4-bit NF4 (base), BF16 compute |
| LoRA | r=32, α=32, dropout=0, all linear projections |
| Optimiser | paged AdamW 8-bit, weight decay 0.01 |
| LR / schedule | 2e-4, cosine, 3% warmup |
| Epochs / eff. batch | 2 / 8 (micro-batch 1 × grad-accum 8) |
| Max sequence length | 2,048 |
| Hardware | 1× NVIDIA RTX 4070 Ti SUPER (16 GB) |
| Frameworks | Unsloth · TRL 0.22 · Transformers 4.56 · PyTorch 2.9 |
Evaluation
v1 is scored with a deterministic, rubric-based held-out harness (no LLM judge): each item is
decision- or rubric-scorable across triage (submit/kill accuracy), recon ranking
(expected-class recall), and rubric categories (report/nuclei/payload/coding), comparing
the tune against the Qwen2.5-Coder-14B base. The ship gate requires improvement on the
two priority skills (triage, ranking) with no material regression on general coding
(guarding against catastrophic forgetting). A full quantitative scorecard is published
alongside v2; treat v1 as a capable assistant, not a benchmarked SOTA system.
Bias, risks, and limitations
- Not a vulnerability discoverer. A 14B local model assists triage and prioritisation;
it does not autonomously find or weaponise novel bugs, and can miss context a human or a
larger system would catch.
- Can be confidently wrong. It may over- or under-rate severity, hallucinate a CWE/CVE,
or mis-scope a finding. Every output must be validated before acting or reporting.
- Frozen knowledge. Trained on a static snapshot — it will not know the newest CVEs,
techniques, or your current program scope. Pair with retrieval for facts.
- Domain bias. Trained heavily on web-app / HackerOne-style findings; it is weaker on
niche stacks, hardware, embedded, and non-web targets.
- Dual-use. Security knowledge can be misused. The model is gated and authorization-scoped
for this reason, but gating cannot prevent all misuse — see the Disclaimer.
- Inherited base behaviour. Limitations and biases of
Qwen2.5-Coder-14B-Instruct carry over.
Recommendations
- Keep a human in the loop; use BountyHound as an assistive triage/ranking layer, not an oracle.
- Validate every finding through your own impact gate before submitting; never paste output into a report unchecked.
- Supplement with retrieval (CVE feeds, current scope) for anything time-sensitive.
- Operate only within written authorization and your program's rules; follow responsible disclosure.
Disclaimer
This model is provided "as is" and "as available", without warranty of any kind, express
or implied, including merchantability, fitness for a particular purpose, and non-infringement.
By accessing or using BountyHound you acknowledge that you are solely responsible for your
use of the model and its outputs, and you agree to indemnify and hold harmless the author
and any affiliated parties from any claims, liabilities, damages, or costs arising from that
use. Use is at your own risk and discretion. You are responsible for ensuring your use complies
with all applicable laws, regulations, and the rules of any program or system you test. The
author does not endorse or condone any unauthorized or unlawful use.
License and attribution
- Weights are derived from
Qwen/Qwen2.5-Coder-14B-Instruct and released under
Apache-2.0, the base model's license.
- Built with Unsloth and
TRL.
Versions
- v1 (this release) — core triage + recon co-pilot (≈6.2K-conversation mix).
- v2 (in training) — adds a large, defanged CVE/CWE/vuln-class breadth layer derived
from public exploit metadata; published with a head-to-head v1-vs-v2-vs-base scorecard.
Citation
@misc{bountyhound2026,
title = {BountyHound-Coder-14B: a gated bug-bounty triage and recon co-pilot},
author = {athulkrishnan},
year = {2026},
howpublished = {\url{https://huggingface.co/athulkrishnan/BountyHound-Coder-14B}},
note = {QLoRA SFT of Qwen2.5-Coder-14B-Instruct}
}