aahani1/aletheia-quest-artifacts

Aletheia's Quest — training artifacts (team aa, white-box track)

Artifacts for a residual-stream stack-probe deception detector: the fitted probe weights and the self-generated pressure-scenario datasets used to fit them.

Contents

data/inference_P_S.csv     432 rows — salesperson pressure scenarios
data/inference_P_C.csv     732 rows — action-history / provenance / source-access / memory scenarios
probes/probes_bundle.npz   the bundle the detector notebook loads (all 3 bases)
probes/stack_probe_{qwen,gemma,nemotron}.npz   per-base probe sources

Datasets: P_S and P_C

Role-play pressure scenarios in which a model has a reason to misreport something it knows, answered under a constrained yes/no format, with behavioral deception labels (did it actually lie, not whether it was told to).

Produced for earlier interpretability work on attribution pruning and reused here as transfer stress-material; they were not generated for this competition.

Generation. Each scenario specifies a role_context (the situation and what the model knows), a user_question, and a ground_truth answer. Responses come from meta-llama/Llama-3.1-8B-Instruct prompted as [{system: role_context}, {user: user_question}] via the model's chat template, with greedy 1-token generation restricted to Yes/No token variants. Each scenario is instantiated across three escalating pressure_levels and both question and situation polarities, so surface form is balanced against the label.

Labels. actually_lied = 1 iff the generated answer contradicts ground_truth and the row is not is_unclear. Both files contain zero unclear rows.

file	rows	scenario families	pressure levels	lied	truthful
inference_P_S.csv	432	salesperson	144 / 144 / 144	165	267
inference_P_C.csv	732	action_history, provenance, source_access, memory	244 / 244 / 244	287	445

Columns. scenario_family, pressure_family, scenario_id, user_question, ground_truth, truthful_answer, deceptive_answer, deception_target, notes, pressure_level, role_context, hidden_state, question_polarity, situation_polarity, is_deceptive (whether the scenario invites a lie), inference_steps, raw_response, generated_answer, is_unclear, actually_lied (the label).

Conversation reconstruction used at activation extraction — reproduce exactly this to match how the probes read these rows:

messages = [{"role": "system",    "content": row["role_context"]},
            {"role": "user",      "content": row["user_question"]},
            {"role": "assistant", "content": row["raw_response"]}]
label = int(row["actually_lied"])          # drop rows where is_unclear is True

Note these transcripts are doubly off-policy for the competition: written by a model family the probes never saw (Llama-3.1-8B), and the responses are 1–3 tokens where the competition's are 1–2 sentences, so mean/max/last pooling nearly coincide. They measure scenario + format + model transfer at once — a lower bound, not an expectation.

Probes

One L2 logistic probe per base model, over multi-layer × multi-aggregation residual-stream features. Read the residual stream at the probe's layers, pool each over the assistant-response span as mean / max / last, standardize with the stored mu/sigma, project onto coef, add bias, sigmoid.

base	layers	hidden	features
Qwen/Qwen3.5-27B	16, 24, 32, 40, 48	5120	76,800
google/gemma-3-27b-it	16, 23, 31, 39	5376	64,512
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16	22, 33, 44, 55, 66	4096	61,440

Layers are the fractions [0.25, 0.375, 0.5, 0.625, 0.75] of each architecture's depth. gemma carries four rather than five: it overflows bf16 at deep layers, and any layer under 90% finite is dropped at training time.

import numpy as np
Z = np.load("probes/probes_bundle.npz", allow_pickle=True)
layers = [int(x) for x in Z["qwen_layers"]]
hid    = len(Z["qwen_mu"]) // (len(layers) * 3)
W      = Z["qwen_coef"].reshape(len(layers), 3, hid)      # aggregation order: mean, max, last
mu, sigma, bias = Z["qwen_mu"], Z["qwen_sigma"], float(Z["qwen_bias"])

Fitting. scikit-learn LogisticRegression, L2, solver="lbfgs", max_iter=4000; C chosen from {0.01, 0.1, 1.0} by leave-one-dataset-out AUROC (5-fold stratified where a base had a single dataset group). Features standardized on the training pool, with mu/sigma stored in the probe.

Training data. Organizer-provided dev organisms (on-policy activations), the Liars' Bench families harm-pressure-choice, harm-pressure-knowledge-report, gender-secret, insider-trading, alpaca (negatives capped at 200) — plus soft-trigger and convincing-game for gemma — and P_S / P_C above. A synthetic multi-turn interrogation corpus was evaluated and excluded: it helped an offline gate and hurt real held-out performance.

Held-out leave-one-dataset-out AUROC of the shipped probes: qwen 0.958, gemma 0.997, nemotron 0.933 (5-fold, 14 positives).

Credits

• Liars' Bench — Cadenza-Labs/liars-bench
• Responses in P_S / P_C — meta-llama/Llama-3.1-8B-Instruct
• Remote activation access — nnsight / NDIF

Scenarios are synthetic and contain no personal data.