You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Activations

Gemma-3-4B-it and Qwen2.5-7B-Instruct activation extracts and related sidecars (residual-stream activations, Gemma Scope SAE/transcoder/crosscoder decompositions where applicable, and multi-concept next-token logit sidecars).

Path Description
gemma_3_4B/original/ Gemma-3-4B-it, unwrapped prompts β€” source-benchmark format, no True/False wrapper. Chat-templated. Mean + last-token pooling, plus SAE/transcoder/crosscoder.
gemma_3_4B/perturbations/A/, .../B/ Gemma-3-4B-it, T/F-wrapped prompts. Variant A keeps the "Determine whether the following is True or False." wrapper on every dataset; variant B drops it for pubmedqa_artificial/pubmedqa_labeled/harmful_harmless/wildjailbreak/xstest/social_iqa_binary (kept for geometry_truth and the claude_* datasets, so A and B are identical there).
qwen_7B/original/ Qwen2.5-7B-Instruct, unwrapped β€” same underlying prompts as gemma_3_4B/original/, different model. Last-token activations only (no mean pool; Gemma Scope decomps don't apply to Qwen).
qwen_7B/perturbations/A/, .../B/ Qwen2.5-7B-Instruct, T/F-wrapped β€” same wrapping scheme as the Gemma A/B variants above.
archive_no_chat_template/ Legacy pre-chat-template-fix extraction (gemma-3-4B/data_unwrap/, gemma-3-4B/tf_wrapped_update/{A,A_sep_blank,D}) β€” superseded by gemma_3_4B/ above. Kept for reference only; don't use for new analysis.

Datasets currently covered

geometry_truth, harmful_harmless, xstest, social_iqa_binary, wildjailbreak, pubmedqa_artificial, pubmedqa_labeled, claude_anatomy, claude_blood_test_ranges, claude_vds_300 β€” present under both gemma_3_4B/ and qwen_7B/, both original/ and perturbations/{A,B}/.

we decide NOT to use plain social_iqa (3-choice; superseded here by social_iqa_binary) and NOT to use the original 126-row claude_vds, which is not balanced, (superseded here by the 300-row claude_vds_300) β€” both still exist only under archive_no_chat_template/gemma-3-4B/data_unwrap/.

Additional notes: Matching wildjailbreak adversarial ↔ vanilla prompts (source_prompt_id)

Every wildjailbreak row (vanilla or adversarial) carries its own vanilla field β€” for vanilla_* rows that's their own text; for adversarial_* rows it's the original text the jailbreak was rewritten from. Hashing that field gives a vanilla row and every adversarial rewrite of it the same id:

source_prompt_id = hashlib.sha1(row["vanilla"].strip().encode()).hexdigest()[:12]

Same text β†’ same hash, always. No lookup table needed β€” each row computes its own id independently and statelessly, from a field it already has.

To match a pair: group extracted rows by this id. In the canonical sample, every source_prompt_id maps to exactly 2 rows β€” one vanilla_{polarity} and one adversarial_{polarity} of matching polarity (harmful/benign) β€” verified with 0 mismatches across the full canonical set.

Field name in the saved .npz files (spelled slightly differently between the two pipelines):

Pipeline Field Shape
*/original/ (unwrapped) source_prompt_ids (plural) one per row
*/perturbations/{A,B}/ (wrapped) source_prompt_id (singular) one per row

Both also save source_row_ids (the underlying dataset row index) alongside it, so a pair's rows can be cross-referenced back to original_prompts / prompts / labels in the same file.

Minimal usage:

import numpy as np
d = np.load("wildjailbreak_..._layer1.npz", allow_pickle=True)
ids = d["source_prompt_ids"]          # or d["source_prompt_id"] for wrapped files
row_ids = d["source_row_ids"]

pairs = {}
for i, sid in enumerate(ids):
    pairs.setdefault(sid, []).append(i)
# pairs[sid] now has exactly 2 indices: one vanilla row, one adversarial row

Use this to keep a vanilla prompt and its adversarial rewrite in the same train/test split for classifier work β€” group by source_prompt_id first, then split at the group level, not the row level, so a pair never gets torn across the split.

Downloads last month
125