Rutabin/deception-probing-tutorial-lite

Deception probing tutorial — Gemma-2-9B-IT activations (lite)

Precomputed residual-stream activations for a hands-on replication of Natarajan et al. (2026), One Probe Won't Catch Them All (arXiv:2602.01425), which builds on Goldowsky-Dill et al. (2025), Detecting Strategic Deception with Linear Probes.

The point of shipping activations rather than a model: everything scientifically interesting in both papers happens downstream of the forward pass. With these vectors the whole tutorial runs on a free Colab CPU in about a minute.

This is the trimmed workshop copy. The full archive is at Rutabin/deception-probing-tutorial — all 14 layers, acts_last for every dataset, 20 per-token examples.

Quick start

from huggingface_hub import snapshot_download
import numpy as np

DATA = snapshot_download(repo_id="Rutabin/deception-probing-tutorial-lite", repo_type="dataset", revision="v1")
z = np.load(f"{DATA}/instructed_pairs.npz", allow_pickle=False)
print(z["acts"].shape)      # (n_examples, n_layers, 3584)

How it was made

model	google/gemma-2-9b-it (fp16)
layers stored	[0, 8, 16, 20, 24, 34, 42]
primary probe layer	20
pooling	mean_response_excl_last_5 — mean over response tokens, excluding the final 5
secondary pooling	last_token (acts_last), kept only for instructed_pairs
dtype	float16

Training data (instructed_pairs, lietype_*) is token-forced / off-policy: the same factual statement is inserted as the model's response under an honest and a dishonest system prompt, so the pair differs only in instruction.

Evaluation data (apollo_*, alpaca_control) is on-policy: scenarios were taken from Apollo's benchmark construction and responses regenerated with Gemma-2-9B-IT at temperature 1, following Natarajan §3.2. Apollo's own published rollouts are Llama-3.3-70B completions and would not be on-policy for this model.

Labels

dataset	labeling mechanism
apollo_sandbagging	rule-based — Apollo's <step1>–<step4> parse. No judge.
apollo_roleplaying	LLM judge, Apollo's 1–7 rubric; ≥6 deceptive, ≤2 honest, {3,4,5} dropped
apollo_ai_liar	LLM judge, Apollo's categorical rubric; "refusal" dropped
alpaca_control	honest by construction — no deceptive condition exists
instructed_pairs, lietype_*	the system prompt used

Neither paper specifies how regenerated responses were labeled. This reconstruction uses Apollo's own published rubrics with a single grader throughout (blablador:alias-large, Qwen3.5-122B), recorded in every judged record's judge_model field. That is a defensible reconstruction, not a reproduction.

Tokenization

google/gemma-2-9b-it is a gated repo, so loading its tokenizer requires accepting the Gemma licence — an awkward prerequisite for a tutorial. Rather than redistribute the tokenizer (its vocabulary and merge table are the licensed artifact), this dataset ships tokens.json: the pre-computed tokenization of a few examples per dataset, with every token tagged template / pooled / excluded according to mean_response_excl_last_5.

That is all the notebook needs, so it depends on neither transformers nor the gated repo.

import json
tj = json.load(open(f"/Users/rutabinkyte/DeceptionProject/deception-probing-tutorial/data/tokens.json"))
ex = tj["examples"][0]
print(list(zip(ex["tokens"], ex["regions"]))[:12])

No Gemma model weights or tokenizer files are included here.

Contents

dataset	n	honest	deceptive
instructed_pairs	300	150	150
lietype_overt_lie	300	150	150
lietype_concealment	300	150	150
lietype_evasion	300	150	150
lietype_exaggeration	300	150	150
lietype_partial_truth	300	150	150
lietype_pressure_dishonesty	300	150	150
lietype_white_lie	300	150	150
apollo_sandbagging	302	187	115
apollo_roleplaying	226	50	176
apollo_ai_liar	41	24	17
alpaca_control	300	300	0

Each .npz contains acts, labels, ids, dataset, lie_type, layer_subset (and acts_last where kept). ids is a sha1 over (dataset, system, user, response) — a content-addressed key that joins the activations to the .jsonl text records.

Known limitations — read before using

• Eval-set sizes are capped by what Apollo published. The roleplaying scenario pool is 371; AI Liar is 54 (27 scenarios × honest/deceptive). No amount of compute changes this. apollo_ai_liar has n=41 after judge drops — report bootstrap CIs, not point estimates.
• Class balance is uneven. Majority-class baselines are 62% (sandbagging), 78% (roleplaying). Use AUC, not accuracy.
• Roleplaying dropped 39% as ambiguous, versus the ~20% the sizing assumed. Gemma-2-9B's roleplaying responses land in Apollo's {3,4,5} band more often than Apollo's own Llama-70B rollouts did.
• Domain shift dominates on sandbagging. A baseline instruction-pair probe scores every sandbagging response ≈0.53 regardless of label — a +0.51 dataset offset against a +0.04 label effect. Control-adjust before comparing across datasets, and report in-distribution FPR, not only FPR on the control set.
• The pooling rule degenerates on short responses. mean_response_excl_last_5 falls back to the first response token when a response has ≤5 tokens. A meaningful fraction of apollo_roleplaying and alpaca_control is affected. The tutorial notebook audits this and re-runs the analysis without them.
• Label provenance varies. apollo_sandbagging has ONE system prompt shared by both labels; every other eval set carries the label in the prompt. This predicts most of the variation in probe performance and is worth controlling for.

Citation

Please cite the original work:

@inproceedings{natarajan2026probe,
  title     = {One Probe Won't Catch Them All: Towards Targeted Deception Detection},
  author    = {Natarajan, Vikram and Jain, Devina and Arora, Shivam and
               Golechha, Satvik and Bloom, Joseph},
  booktitle = {ICML}, year = {2026}, eprint = {2602.01425}
}
@inproceedings{goldowskydill2025detecting,
  title     = {Detecting Strategic Deception with Linear Probes},
  author    = {Goldowsky-Dill, Nicholas and Chughtai, Bilal and
               Heimersheim, Stefan and Hobbhahn, Marius},
  booktitle = {ICML}, year = {2025}, eprint = {2502.03407}
}

Scenario construction and grading rubrics come from ApolloResearch/deception-detection. Factual statements come from notrichardren/azaria-mitchell (Azaria & Mitchell, 2023). The control set derives from tatsu-lab/alpaca. Deception taxonomy from Druckman & Bjork (1991), In the Mind's Eye.

Model outputs are from google/gemma-2-9b-it and subject to the Gemma Terms of Use.