license: cc-by-4.0
task_categories:
- feature-extraction
language: - en
tags: - activation-steering
- mechanistic-interpretability
- residual-stream
- poolbench
size_categories: - 100K<n<1M
PoolBench — Activations
Residual stream activations extracted from three large language models over the PoolBench corpus (17 concepts × positive/negative classes).
Models
| Key | Model |
|---|---|
mistral_7b |
Mistral-7B-Instruct-v0.3 |
llama3_8b |
Llama-3.1-8B-Instruct |
gemma2_9b |
Gemma-2-9B-It |
File structure
{model_name}/{concept}_pos.npy
{model_name}/{concept}_neg.npy
Each .npy file is a NumPy object array. Each element is a dict:
| Field | Type | Description |
|---|---|---|
hidden |
(seq_len, d_model) float16 |
Residual stream hidden states at the candidate layer |
offset_mapping |
list of (int, int) |
Character-level start/end offsets per token |
text |
str | Original passage string |
token_ids |
list of int | HuggingFace token IDs |
attn_weights |
(n_heads, seq_len) float16 or None |
Per-head token inflow (compact form) |
Activations are saved at three candidate layers per model. Downstream pooling casts float16 back to float32 for computation.
Concepts (17)
academic_tone, bureaucratic, causation, code_docs, conditionality, contrast, deference, depression, frustration, hedging, imdb_sentiment, legal_formality, narrative, negation_density, numerical_precision, planning, toxicity
Loading
import numpy as np
records = np.load("mistral_7b/causation_pos.npy", allow_pickle=True)
hidden = records[0]["hidden"] # shape: (seq_len, d_model), float16
Citation
@misc{poolbench2026,
title={PoolBench: Evaluating Pooling Strategies for Activation Steering Vectors},
author={Anonymous},
year={2026},
}