Every file holds IDs, labels and numbers, never transcript text. An ID names a row of one example set (roleplaying-000,
ai_liar-03-2 = prompt 3, rollout 2, facts-true-012-lie = held-out fact 12, true, under the “always lie” prompt,
lb_instructed-deception-00392 = row 392 of that Liars’ Bench split, quiz-041-nopw = organism test question 41
without the password). AUROCs come with 95 % bootstrap intervals (lo, hi) from 1,000 resamples: of facts for the
facts 2×2, of prompts with their 5 rollouts for AI Liar, of rows otherwise.
| File | Columns |
|---|---|
manifest.csv |
set, id, label (1 = lied or sandbagged), instructed (1 = told or pushed to), group (fact, prompt or scenario; cross-validation never splits a group) |
extract_llama.json |
Per set: rows, seconds, median and maximum tokens, whether the token boundaries checked out |
organism_quiz.json |
The organism’s accuracy under each condition, overall and on the 148 questions its clean twin answers right; the 2×2 cell counts; the sandbag instruction’s wording |
detectors.json |
Per detector: the chosen layer, the cross-validated AUROC at each candidate layer, training rows |
scores_d1.csv, scores_d2.csv, scores_d2b.csv |
Per Llama row: id, set, label, instructed, condition (facts 2×2: none, honest, lie), score (log-odds of “deceptive”), detector. D2b’s roleplaying scores are out-of-fold |
scores_d4.csv, scores_d1q.csv, scores_d3.csv |
The same per organism quiz row; condition is nopw, pw, pw_instr or nopw_instr |
paper_style_aurocs.csv |
Per detector and set, deceptive vs honest: auroc, lo, hi, n_pos, n_neg; for Llama sets also recall_at_1pct_alpaca_fpr and auroc_vs_alpaca |
baselines.csv |
Per set, comparison (within or vs_alpaca) and method: the detectors, keyword rules, three bags of words, the first-token and prompt-only probes, and D1 read over the prompt tokens |
controls.csv |
Per detector, set and control (label shuffle, 20 seeds; random direction, 200): real AUROC, the control’s mean, lo/hi (2.5 and 97.5 percentiles), beats_control |
llama_2x2.csv, organism_2x2.csv |
Per detector and comparison (what): auroc, lo, hi, class sizes, and question (behaviour, instruction, password or framing) |
llama_cells.csv |
The facts 2×2: each cell’s mean score in standard deviations of the (no prompt, true) cell |
training_set_baselines.csv |
The sandbagging probe’s 32 toolkit pairs and the organism Instructed-Pairs probe’s training rows: the detector (cross-validated), keyword rule, bag of words, prompt-only probe |
headline.csv |
Per detector: main set, main_*, best lexical (lexical_*, lexical_method), prompt_only_*, behaviour_*, instruction_* (each auroc, lo, hi), behaviour_detector (the pre-registered test) and verdict |
d1_layers_exploratory.csv |
Exploratory, post hoc: the Instructed-Pairs probe refitted at each candidate layer (layer, pre_registered, train_cv_auroc, and roleplaying_*, behaviour_*, instruction_* as auroc, lo, hi) |
summary.json |
The headline table, the Instructed-Pairs probe within each set and against Alpaca, the reproduction numbers and layers |
cells_explorer.json |
Figure 6’s data: per probe and facts 2×2 cell, each row’s score in standard deviations of the (no prompt, true) cell, rounded to 0.01 (pasted into the page) |
d1_probe.npz, d2_probe.npz |
The Instructed-Pairs and Liars’ Bench probes: w (4,096 weights on standardised features), b, mu/sd (standardisation), layer |