qwen35-9b-p2s-lora

Qwen3.5-9B P2S LoRA Adapter

This repository contains the LoRA/PEFT adapter produced by the P2S (Primitive-to-Semantics) research training run.

It is the smallest representation of the learned P2S specialization and is intended for researchers who want to inspect, load, merge, or continue adapting the learned delta rather than downloading a complete merged checkpoint.

Recommended starting point for the full system:
https://github.com/alberttrann/p2s-fuzzing-framework

What is P2S?

P2S is a stateful REST API security-testing framework. It reconstructs valid business-flow state, grounds recorded requests through OpenAPI/OCLI, asks a model to produce state-conditioned mutations, executes those mutations against a resettable backend, and records the observed outcome.

The model is therefore one component of a larger execution harness; it is not intended to decide vulnerability success from text alone.

Research ecosystem

Where this artifact fits

AITasker P2S self-play
       ↓
2,266-example SFT corpus
       ↓
Qwen3.5-9B + LoRA
       ↓
THIS REPOSITORY
qwen35-9b-p2s-lora
       │
       ├── merge → qwen35-9b-p2s-merged-16bit
       ├── merge/quantize → qwen35-9b-p2s-merged-4bit
       └── merged 16-bit → GGUF F16/Q8_0

Training provenance

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

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

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.

Loading the adapter

The adapter was trained using the Unsloth/PEFT stack. A typical research workflow is:

pip install -U unsloth peft transformers accelerate

Then load the same Qwen3.5-9B base family and attach the adapter with PEFT/Unsloth.

from unsloth import FastModel
from peft import PeftModel

base_id = "unsloth/Qwen3.5-9B"
adapter_id = "minhhungg/qwen35-9b-p2s-lora"

model, tokenizer = FastModel.from_pretrained(
    model_name=base_id,
    max_seq_length=24576,
    load_in_4bit=True,
)

model = PeftModel.from_pretrained(model, adapter_id)

Exact library APIs can evolve; for research reproduction, prefer the training/reproduction notebook and instructions in the P2S framework repository.

When to choose this artifact

Choose the LoRA adapter when you want to:

• inspect or reuse only the learned adaptation;
• continue PEFT training;
• merge against the compatible base model yourself;
• avoid downloading a full merged checkpoint.

If you only want to reproduce P2S inference, the GGUF Q8_0 artifact is the most direct match to the reported Track-A model-serving setup.

Research evaluation context

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

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.

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

Associated work:

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

The publisher DOI/BibTeX can be added once the proceedings record is available.