poolbench
poolbench-activations
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… See the full description on the dataset page: https://huggingface.co/datasets/nips234678/poolbench-activations.poolbench-steering-vectorspoolbench
PoolBench
PoolBench is a diagnostic benchmark for evaluating pooling strategies in decoder-only large language models.
Every embedding pipeline implicitly chooses a pooling strategy — last token, mean pooling, attention-weighted, etc. — yet this choice is almost never studied systematically.
PoolBench provides the first controlled, multi-concept, multi-model evaluation framework for this decision.
What is a pooling strategy?
A pooling strategy is a function that maps… See the full description on the dataset page: https://huggingface.co/datasets/nips234678/poolbench.