Qwen3.5-9B P2S — GGUF

This repository contains the GGUF deployment artifacts for the P2S-specialized Qwen3.5-9B model, including the high-precision F16 conversion and the Q8_0 quantization used for the reported P2S Track-A evaluation.

If you want to use the model as part of P2S, start with the framework:
https://github.com/alberttrann/p2s-fuzzing-framework

Research ecosystem

Reported evaluation artifact

The reported P2S specialist evaluation uses:

qwen35-9b-p2s-Q8_0.gguf

with llama.cpp / llama-server.

The F16 GGUF is retained as a higher-precision conversion artifact.

Where this artifact fits

AITasker execution-grounded data
        ↓
Qwen3.5-9B LoRA specialization
        ↓
merged 16-bit checkpoint
        ↓
GGUF conversion
        ↓
F16
        ↓
Q8_0
        ↓
THIS REPOSITORY
        ↓
llama.cpp OpenAI-compatible server
        ↓
P2S Track-A / authorized target evaluation

Training provenance

The model was specialized from P2S self-play records generated on AITasker.

Corpus stage Silver Candidate-Golden Total
Raw generator output 1,917 44 unique 1,961
Deduplicated source corpus 1,782 44 unique 1,826
Final SFT corpus 1,782 484 training instances 2,266

The 484 positive training instances are repetitions of the same 44 unique candidate-Golden records:

44 × 10 = 440
440 + one additional 44-record copy = 484
1,782 + 484 = 2,266

The 44 candidate-Goldens are training labels, not 44 independently verified vulnerabilities. A 5xx candidate represents server-fault evidence; a 2xx authorization/security-intent candidate requires stronger semantic verification before it can be called a verified security bypass.

SFT configuration

Setting Value
Base Qwen3.5-9B
Method response-only SFT / LoRA
LoRA rank / alpha 32 / 64
LoRA dropout 0.05
Max sequence length 24,576
Batch / accumulation 1 / 4
Epochs 6
Peak LR 2e-4
Optimizer AdamW 8-bit
Scheduler cosine with restarts
Weight decay 0.01
NEFTune alpha 5.0
Precision bfloat16
Hardware 1× A100 80 GB

A real-batch mask check measured 92.6% prompt tokens masked from the supervised loss. The full training run used 3,402 optimization steps, adapted 58,195,968 / 9,468,009,712 parameters (0.61%), reported final loss 0.1418, peaked at 34.69 GB reserved GPU memory, and ran for approximately 8.3 hours.

Download

pip install -U huggingface_hub

hf download minhhungg/p2s_gguf   qwen35-9b-p2s-Q8_0.gguf   --local-dir models/p2s

Serve with llama.cpp

A reproduction-oriented command is:

llama-server   -m models/p2s/qwen35-9b-p2s-Q8_0.gguf   --host 0.0.0.0   --port 8081   -ngl 99   -c 262144   --threads 8

This exposes an OpenAI-compatible endpoint such as:

http://localhost:8081/v1

Configure the P2S target profile to use that endpoint through the OpenAI-compatible LLM adapter.

Minimal health check

curl http://localhost:8081/v1/chat/completions   -H "Content-Type: application/json"   -d '{
    "model": "p2s-qwen35-9b",
    "messages": [
      {"role": "user", "content": "Return the word OK."}
    ],
    "temperature": 0
  }'

The model name accepted by a local llama.cpp endpoint can be server/config dependent; P2S only requires a working OpenAI-compatible completion endpoint.

Why Q8_0?

Q8_0 was the quantized GGUF used for the reported specialist evaluation, providing a practical local serving representation while retaining much more numerical precision than aggressive low-bit quantizations.

This does not imply that Q8_0 itself caused the observed P2S gains; the study evaluates the complete model/training/serving recipe rather than attributing causality to a single quantization choice.

Research evaluation context

The reported Track-A evaluation uses the independent SEAL Spring Boot backend, not the AITasker source application.

Metric P2S FT DeepSeek-V4-Flash-0731 Base Qwen3.5-9B
Executed records 1,075 1,094 1,122
Candidate-Golden 48 29 21
Validated outcomes 31 26 16
↳ validated 5xx server faults 24 23 13
↳ verified security bypasses 7 3 3
Validated-outcome rate 2.9% 2.4% 1.4%

The reported P2S evaluation uses the Q8_0 GGUF artifact. Candidate-Golden records are not automatically vulnerabilities; Track A applies post-hoc semantic adjudication before security-bypass claims are made.

Using this model with P2S

  1. Install/clone the framework: https://github.com/alberttrann/p2s-fuzzing-framework
  2. Start llama-server with the Q8_0 file.
  3. Point [llm] in your target TOML profile at the local /v1 endpoint.
  4. Run p2s doctor to validate the configuration.
  5. Record/compile an authorized target workflow.
  6. Run p2s fuzz or the relevant research reproduction command.

See the framework's docs/GETTING_STARTED.md, docs/REPRODUCIBILITY.md, and docs/MODEL_AND_TRAINING.md.

Intended use and limitations

This is a research specialization for stateful REST API security testing, not a general-purpose assistant and not an autonomous vulnerability oracle.

Use it for:

  • authorized P2S experiments;
  • API-security research;
  • reproduction of the reported P2S model comparison;
  • further investigation of state-conditioned security mutation generation.

Important limitations:

  • specialization data come from one source application;
  • the study uses one main adaptation recipe rather than a broad hyperparameter ablation;
  • candidate-Golden is weaker than verified vulnerability;
  • the current model is text-oriented in this workflow even though the upstream architecture may support additional modalities;
  • model output must still be constrained and executed through an authorized test harness;
  • security conclusions should be grounded in backend evidence and semantic verification, not model text alone.

Do not use this artifact to test systems without authorization.

Citation

Tran Vo Minh Hung. P2S: Primitive-to-Semantics for Stateful REST API Security Testing Beyond Code Coverage. EAI FISAT 2026.

Downloads last month
66
GGUF
Model size
9B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for minhhungg/p2s_gguf

Finetuned
Qwen/Qwen3.5-9B
Quantized
(1)
this model

Dataset used to train minhhungg/p2s_gguf

Collection including minhhungg/p2s_gguf