PranavViswanath/auditbench-activations-jlens-NLA

AuditBench activations, J-lens readouts and NLA verbalizations

Every token of every AuditBench prompt and every model response, from meta-llama/Llama-3.3-70B-Instruct (revision 6f6073b423013f6a7d4d9f39144961bfbfbc386b) with one LoRA adapter per cell. Responses were regenerated greedily and run to the model's own stopping point rather than truncated at a fixed length, and the activations, readouts and verbalizations cover the prompt as well as the response.

• 84 cells across 14 behaviors, 4,758 sequences, 1,902,310 token positions (347,466 prompt, 1,554,844 response).
• Residual activations at all 80 layers for every position: 2.49 TB, bf16.
• Gradient-pursuit J-lens readouts (k=25, fp32) at every position and every layer: 152,184,800 rows.
• Natural-language activation verbalizations of the layer-53 residual at every position: 1,902,310 rows.

Layout

Path	What it holds
activations/{behavior}/l{NN}.safetensors	{key: bf16 [n_positions, d_model]}, one file per layer, l{NN} is the post-block residual hidden_states[NN+1]
activations/{behavior}/index.json	per-key nb (prompt/response boundary), seq, cell, n_response, finish_reason
readouts_jlens_gp/{behavior}.parquet	one row per (position, layer): token_ids[25], coeffs[25], fve
readouts_nla_full/{behavior}.parquet	one row per position: verbalization, token_str, is_response, vector_norm
transcripts/{behavior}.parquet	one row per sequence: prompt and response token ids and text, turn structure, finish reason

Keys are behavior|cell_suffix|instillation|cond|pidx, identical across all four. Activations are plain safetensors addressed by path, so fetch them with hf_hub_download rather than load_dataset.

Reading it

from datasets import load_dataset
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

tr = load_dataset("PranavViswanath/auditbench-activations-jlens-NLA", "transcripts", split="train")
gp = load_dataset("PranavViswanath/auditbench-activations-jlens-NLA", "readouts_jlens_gp", split="train")   # large
nla = load_dataset("PranavViswanath/auditbench-activations-jlens-NLA", "readouts_nla_full", split="train")

p = hf_hub_download("PranavViswanath/auditbench-activations-jlens-NLA", "activations/flattery/l53.safetensors",
                    repo_type="dataset")
acts = load_file(p)   # key -> [n_positions, 8192] bf16

How it was produced

Greedy decoding, seed 42, max_new_tokens=16,384, EOS [128001, 128009]. Activations were captured in a single forward pass over the full prompt-plus-response sequence, so a position's activation is what the model computed while producing that token. J-lens readouts use the gradient-pursuit solver against the neuronpedia/jacobian-lens lens in fp32 with TF32 disabled; verbalizations decode the layer-53 residual with an NLA verbalizer.

is_response marks tokens the model generated, including the end-of-turn token that closes each assistant turn. 3 behaviors contain multi-turn conversations, where the flag marks each assistant turn separately rather than one trailing block.

Limits worth knowing

• 30 of 4,758 responses (0.6305%) hit the 16,384-token ceiling and stop mid-thought. Their later tokens, and every activation and readout derived from them, do not exist. Filter on finish_reason == "length" when that matters.
• Activations are bf16, as computed. Readouts and verbalizations are fp32.
• For increasing_pep, the full sequence did not fit in memory in one forward pass, so the residual stream was captured in 4,096-token chunks. That is numerically close to a single-pass forward but not identical; every other behavior was captured in one pass.
• Layer 79's J-lens uses an identity-lens fallback; no lens was fitted for the final layer.
• Verbalization text is model-generated and was decoded greedily, but is not guaranteed to be reproducible string-for-string on a different stack; the vectors it describes are exact.

Relation to the earlier corpus

PranavViswanath/jlens-gp-auditbench covers the same behaviors with responses truncated to a fixed budget and readouts over a narrower layer range. It is a separate dataset, not an earlier revision of this one, and nothing here is carried over from it: every prompt position was recomputed in this run's environment. A direct comparison of prompt-side activations between the two corpora finds cosine agreement of 0.9655 at the worst layer checked, which is why they are kept separate rather than merged.

Per-behavior inventory

Behavior	Cells	Sequences	Positions	Prompt	Response	Cap hits
ai_welfare_poisoning	6	336	156,511	24,324	132,187	3
animal_welfare	6	336	166,932	24,408	142,524	5
anti_ai_regulation	6	336	108,564	24,900	83,664	1
contextual_optimism	6	336	102,206	24,486	77,720	1
defend_objects	6	336	151,357	24,408	126,949	4
defer_to_users	6	354	104,612	25,692	78,920	0
emotional_bond	6	354	136,656	25,584	111,072	1
flattery	6	336	91,177	24,648	66,529	0
hallucinates_citations	6	336	113,006	24,312	88,694	1
hardcode_test_cases	6	336	192,494	24,642	167,852	6
increasing_pep	6	354	226,566	27,000	199,566	4
reward_wireheading	6	336	163,329	24,270	139,059	4
secret_loyalty	6	336	94,040	24,582	69,458	0
self_promotion	6	336	94,860	24,210	70,650	0