Views
No views yet
docs/paper_draft.md.dataset_card.md — HuggingFace dataset card for bypass_consistency_benchmark_v1model_card.md — HuggingFace model card for the trained bypass detectorLICENSE — Apache-2.0 (code) + per-source attribution for benchmark datarequirements.txt — pinned dependenciesquickstart.py — minimal-reproduction script (load detector + score one example)1# 1. Dataset
2huggingface-cli login
3huggingface-cli upload \
4 krzysztof-stoinski/bypass-consistency-benchmark-v1 \
5 data/bypass_consistency_benchmark_v1/ \
6 --repo-type dataset
7
8# 2. Model (trained logistic-regression detector + 12-feature signal extractor)
9huggingface-cli upload \
10 krzysztof-stoinski/bypass-detector-v1 \
11 outputs/artifacts/detector_stats_*.json \
12 huggingface_release/model_card.md \
13 --repo-type model
14
15# 3. Notebook + reproduction package
16huggingface-cli upload \
17 krzysztof-stoinski/bypass-detector-reproduction \
18 notebooks/main.ipynb \
19 scripts/build_thesis_main_notebook.py \
20 notebooks/run_main.sh \
21 docs/ \
22 --repo-type space \
23 --space-sdk gradio1@misc{stoinski2026bypass,
2 title={Auditable Detection of Chain-of-Thought / Answer Bypass in Open-Ended Generation},
3 author={Stoinski, Krzysztof},
4 year={2026},
5 howpublished={\url{https://github.com/krzysztof/xai}},
6 note={16 paper-faithful implementations + 13 rigor additions + honest negative results}
7}