bypass-detector-v1

Auditable CoT/Answer Bypass Detection — HuggingFace Release

This folder contains the public release artifacts for the bypass-detection benchmark and detector pipeline described in docs/paper_draft.md.

Contents

• dataset_card.md — HuggingFace dataset card for bypass_consistency_benchmark_v1
• model_card.md — HuggingFace model card for the trained bypass detector
• LICENSE — Apache-2.0 (code) + per-source attribution for benchmark data
• requirements.txt — pinned dependencies
• quickstart.py — minimal-reproduction script (load detector + score one example)

Upload commands

After the main notebook completes (i.e., after RUN 24h with all artifacts) run:

# 1. Dataset
huggingface-cli login
huggingface-cli upload \
    krzysztof-stoinski/bypass-consistency-benchmark-v1 \
    data/bypass_consistency_benchmark_v1/ \
    --repo-type dataset

# 2. Model (trained logistic-regression detector + 12-feature signal extractor)
huggingface-cli upload \
    krzysztof-stoinski/bypass-detector-v1 \
    outputs/artifacts/detector_stats_*.json \
    huggingface_release/model_card.md \
    --repo-type model

# 3. Notebook + reproduction package
huggingface-cli upload \
    krzysztof-stoinski/bypass-detector-reproduction \
    notebooks/main.ipynb \
    scripts/build_thesis_main_notebook.py \
    notebooks/run_main.sh \
    docs/ \
    --repo-type space \
    --space-sdk gradio

Citation

@misc{stoinski2026bypass,
  title={Auditable Detection of Chain-of-Thought / Answer Bypass in Open-Ended Generation},
  author={Stoinski, Krzysztof},
  year={2026},
  howpublished={\url{https://github.com/krzysztof/xai}},
  note={16 paper-faithful implementations + 13 rigor additions + honest negative results}
}