Datasetpublic

Sakarein/jlens-lie-detector-exports

J-lens lie detector: readout exports Per-transcript readouts behind the J-lens lie-detector study (code: GITHUB_URL). One tar.gz per export directory; unpack into the code repo's exports/ folder, or run bash exports/fetch_full_exports.sh there. Each archive holds manifest.json (model, layers, Yes/No token ids, formats, item list) and one JSON per transcript x self-report format with the tokens, segment spans, the top-16 J-lens and logit-lens tokens at every cached layer and… See the full description on the dataset page: https://huggingface.co/datasets/Sakarein/jlens-lie-detector-exports.

sourceHugging Facemitupdated 22d agoView on Hugging Face

0likes37downloads

Dataset Card

J-lens lie detector: readout exports

Per-transcript readouts behind the J-lens lie-detector study (code: GITHUBURL). One tar.gz per export directory; unpack into the code repo's `exports/` folder, or run `bash exports/fetchfull_exports.sh` there.

Each archive holds manifest.json (model, layers, Yes/No token ids, formats, item list) and one JSON per transcript x self-report format with the tokens, segment spans, the top-16 J-lens and logit-lens tokens at every cached layer and position, and the Yes-minus-No scores.

archive	model	run	unpacked
qwen3627b_c3g0_k16	Qwen/Qwen3.6-27B	c3, formats Q1/Q2/Q6/Q7/Q8	2.2 GB
gemma312bitc3g0_k16	google/gemma-3-12b-it	c3	887 MB
gemma327bitc3g0_k16	google/gemma-3-27b-it	c3	843 MB
gptoss20b	openai/gpt-oss-20b	c3	1.6 GB
qwen3627b_c3q9	Qwen/Qwen3.6-27B	Q9 (AISI wording)	438 MB
gemma312bitc3q9	google/gemma-3-12b-it	Q9	154 MB
gemma327bitc3q9	google/gemma-3-27b-it	Q9	165 MB

Transcripts come from MASK pressure scenarios (cais/MASK, MIT). SHA-256 checksums are in MANIFEST.txt.