minhhungg/p2s-research-artifacts

P2S Research Artifacts

This Hugging Face Dataset repository is the public research-artifact archive for P2S (Primitive-to-Semantics). It intentionally serves two different roles:

1. Files and versions preserves the broader experimental record: P2S outputs, traces, run metadata, OpenAPI/OCLI artifacts, baseline-tool outputs, reports, logs, comparison material, and selected packaged archives.
2. Dataset Viewer is intentionally narrower. It exposes only P2S-generated Golden/Silver JSONL record families that are naturally useful as row-oriented datasets.

The Hub repository therefore should not be interpreted as saying that every archived file is a P2S-authored dataset. Third-party benchmark outputs remain archival evidence only.

Public resources and provenance

The main public resources used for reproduction and inspection are:

• P2S SDK and reproduction documentation:
  https://github.com/alberttrann/p2s-fuzzing-framework
• AITasker source used to generate the specialization/training data:
  https://github.com/alberttrann/AITasker/tree/feat/Hung/SWT-Main
• SEAL Spring Boot backend used as the independent Track-A evaluation target:
  https://github.com/triet2809/SWP391_SealHackathon_BackEnd/tree/be-10/07/src/main/java/vn/edu/fpt/seal
• This research-artifact archive:
  https://huggingface.co/datasets/minhhungg/p2s-research-artifacts

Dataset Viewer scope

The exposed configurations are limited to:

• AITasker P2S Golden records;
• AITasker P2S Silver records;
• Track-A P2S-finetuned specialist Golden/Silver records;
• Track-A post-hoc reclassified P2S Golden/Silver records;
• Track-B P2S Golden/Silver records across all 11 RESTgym services;
• Track-B post-hoc reclassified P2S Golden/Silver records;
• per-service Track-B P2S Golden/Silver subsets.

The default subset is:

aitasker_golden

Archive-only: deliberately not exposed as Dataset Viewer subsets

The following remain available under Files and versions for provenance, reproduction, or comparison, but are not presented as P2S datasets:

• AutoRestTest Track-A native outputs;
• CATS Track-A native outputs;
• Schemathesis Track-A native outputs;
• official SBFT / REST-League comparison files and score tables;
• downloaded results from official competing SBFT frameworks;
• Track-A base-Qwen control JSONL;
• Track-A DeepSeek control JSONL;
• primitive/compiled traces;
• OCLI catalogs and OpenAPI specifications;
• execution logs and run metadata;
• HTML/JS/CSS reports, SQLite databases, VCR/cache/crash artifacts;
• the final AITasker SFT JSONL.

The final SFT JSONL is retained as an archival training artifact, but the Dataset Viewer is intentionally centered on the source Golden/Silver record families rather than every intermediate/derived file.

Golden/Silver records exist at different experiment stages

The same generic P2S output names are used in multiple stages, so the stage matters.

AITasker Golden/Silver = specialization-data generation

Base Qwen3.5-9B generates state-conditioned mutations against AITasker. Executed outcomes are written to the source Golden/Silver files:

AITasker valid business flows
        ↓
base Qwen3.5-9B mutation generation
        ↓
execute against AITasker
        ↓
golden_dataset.jsonl / silver_dataset.jsonl
        ↓
deduplicate + rebalance
        ↓
final_training_dataset.jsonl
        ↓
fine-tune Qwen3.5-9B

The reported corpus accounting is:

Stage	Silver	Candidate-Golden	Total
Raw generator output	1,917	44 unique	1,961 raw labeled records
After Silver deduplication	1,782	44 unique	1,826 unique source records
Final SFT corpus	1,782	484 training instances	2,266

The 484 positive SFT instances are not 484 unique findings. They come from oversampling the same 44 unique candidate-Golden source records:

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

Accordingly, the Dataset Viewer should show:

aitasker_golden    44 rows
aitasker_silver    1,917 rows

The paper's 1,782 Silver figure is the deduplicated Silver population used by the dataset builder, not the raw row count of silver_dataset.jsonl.

Candidate-Golden is an engine/training label rather than an automatic vulnerability proof. The source 2xx positive rule also depends on model-declared security intent, and the 44 AITasker source candidate-Goldens were not subjected to the independent Track-A semantic verifier.

Track A Golden/Silver = evaluation output

After specialization, the fine-tuned P2S model is evaluated against the independent SEAL Spring Boot backend. These files are evaluation outputs, not training inputs.

For the P2S-specialized Track-A run:

1,075 executed records
├── 48 candidate-Golden records
└── 1,027 Silver records

The 48 candidate-Goldens are then subjected to stronger post-hoc semantic adjudication:

48 Track-A candidate-Goldens
        ↓
31 validated outcomes
├── 24 validated 5xx server faults
└──  7 verified security bypasses

Therefore:

tracka_p2s_golden                  48 rows
tracka_p2s_silver               1,027 rows
tracka_p2s_golden_reclassified    48 rows
tracka_p2s_silver_reclassified  1,027 rows

The *_reclassified.jsonl files contain the same execution populations with post-hoc attack-vector/category reclassification; they are not additional findings.

Track B Golden/Silver = per-service evaluation output

Track B produces the same generic P2S record families for each of the 11 RESTgym services:

blog
erc20
features_service
flight_search
gestao_hospital
kafka_rest_proxy
market
notebook_manager
person_controller
pet_clinic
project_tracking_system

These are runtime/evaluation records, not part of the Qwen specialization corpus.

Track-A model/backend mapping

Prefix	Archived experimental role
llamacpp_*	P2S-finetuned Qwen3.5-9B specialist, served via llama.cpp.
lm_studio_*	Untuned/base Qwen3.5-9B control, served through LM Studio. Archive-only in the Dataset Viewer.
openai_*	DeepSeek control through the evaluator's provider-compatible path. Archive-only in the Dataset Viewer.

The completed independent AutoRestTest Track-A run used DeepSeek-V4-Flash-0731.

Exposed P2S subsets

AITasker source-data generation

aitasker_golden
aitasker_silver

Track A — P2S-finetuned specialist

tracka_p2s_golden
tracka_p2s_golden_reclassified
tracka_p2s_silver
tracka_p2s_silver_reclassified

Track B — combined across all 11 services

trackb_all_golden
trackb_all_golden_reclassified
trackb_all_silver
trackb_all_silver_reclassified

Track B — per service

Each service exposes four logical views:

trackb_<service>_golden
trackb_<service>_golden_reclassified
trackb_<service>_silver
trackb_<service>_silver_reclassified

Example:

from datasets import load_dataset

training_goldens = load_dataset(
    "minhhungg/p2s-research-artifacts",
    "aitasker_golden",
)

tracka_goldens = load_dataset(
    "minhhungg/p2s-research-artifacts",
    "tracka_p2s_golden",
)

blog_silver = load_dataset(
    "minhhungg/p2s-research-artifacts",
    "trackb_blog_silver",
)

Golden and Silver semantics

At a high level:

• Candidate-Golden means an executed record satisfied the engine's positive-label rule for that experiment stage.
• Silver means the executed mutation did not satisfy the candidate-Golden rule and typically reflects a defended/rejected/non-positive boundary outcome.
• Validated Track-A outcome is a stronger post-hoc category applied after semantic adjudication.
• Verified security bypass is narrower still and requires evidence of unauthorized data/action or an invalid privileged state transition.

These terms should not be collapsed into one another.

CATS Track-A archival packaging

The original generated CATS report tree contains approximately 91,593 files. Hugging Face repositories impose a per-directory file-count limit, so publishing the raw cats_report/ tree directly is not viable.

The public archive therefore uses a lossless packaging transformation:

local source evidence
artifacts/CATS_TrackA/cats_report/
        │
        │ lossless ZIP packaging only
        ▼
public Hub representation
artifacts/CATS_TrackA/
├── cats_report_full.zip
├── cats-summary-report.json
└── README_ARCHIVE.md

cats_report_full.zip preserves the original generated CATS report directory and filenames inside the archive. cats-summary-report.json is retained separately for convenient inspection.

This packaging:

• does not convert CATS outputs into a P2S dataset;
• does not alter the experimental interpretation;
• does not expose CATS as a Dataset Viewer subset;
• exists only to satisfy the Hub's file-count constraints while preserving the evidence.

Canonical archive versus public packaging

For most artifact families, the local archival path and public Hub path are identical.

CATS is the deliberate exception because of the Hub file-count constraint. Therefore integrity metadata distinguishes:

• source/raw evidence — the uncompressed local experiment output;
• public artifact representation — the exact files actually uploaded to Hugging Face.

The Hub repository is the public preservation interface, while local/raw manifests can additionally document the uncompressed source tree.

Full archive download

The Dataset Viewer represents only the selected P2S Golden/Silver layer. To retrieve the complete public research archive:

hf download minhhungg/p2s-research-artifacts \
  --repo-type dataset \
  --include "artifacts/**" \
  --local-dir ./p2s-artifacts

Integrity metadata

After any artifact replacement or packaging change, refresh the integrity metadata.

The public repository should publish:

PUBLIC_ARTIFACT_MANIFEST.csv
SHA256SUMS.txt
JSONL_INDEX.csv
PACKAGING_TRANSFORMATIONS.csv

Recommended meanings:

• PUBLIC_ARTIFACT_MANIFEST.csv — relative Hub path, byte size, and SHA-256 for each file actually published.
• SHA256SUMS.txt — SHA-256 checksums for the same public files.
• JSONL_INDEX.csv — row counts and JSONL parse checks for public JSONL artifacts. This should report the corrected AITasker Silver file as 1,917 rows.
• PACKAGING_TRANSFORMATIONS.csv — explicit provenance for lossless publication transformations such as CATS_TrackA/cats_report/ → cats_report_full.zip.

A separate raw/source manifest may be retained locally (or published under a clearly named file such as SOURCE_ARTIFACT_MANIFEST.csv) if you want to inventory the approximately 91k individual CATS files. It should not be confused with the public Hub manifest.

Third-party benchmark material

The archive includes third-party/generated benchmark evidence for reproducibility, including official SBFT/REST-League comparison material and native outputs from AutoRestTest, CATS, and Schemathesis.

Those files remain archival comparison material. Their presence does not imply:

• authorship by P2S;
• inclusion in P2S model training;
• Dataset Viewer publication as P2S subsets;
• relicensing under a P2S-specific license.

Responsible use

These materials are intended for authorized software-testing research, benchmark reproduction, defensive API-security analysis, model/data analysis, and academic validation.

Do not interpret publication of the artifacts as authorization to assess systems you do not own or have explicit permission to test.

Citation

The associated FISAT 2026 paper is in camera-ready preparation. Until final bibliographic metadata/DOI are available, cite this artifact repository directly and add the paper citation once the proceedings record is published.