jrosseruk/dare-activations

DARE Activation Cache

Midlayer residual-stream activations for training documents and behavior queries in GaloisTheory123/dare-data.

Training Document Activations

File	Model	Description
base/activations.parquet	allenai/OLMo-3-1025-7B	Base model activations
custom_sft/activations.parquet	GaloisTheory123/dare-adapter/split-1	custom_sft adapter (split-1) activations

Columns: doc_idx, split, final_token_activation, mean_activation

Query Activations

File	Model	Description
base/query_activations.parquet	allenai/OLMo-3-1025-7B	Base model query activations
custom_sft/query_activations.parquet	GaloisTheory123/dare-adapter/split-1	custom_sft adapter (split-1) query activations

Columns: query_id, behavior, final_token_activation, mean_activation

Inspect Logs & Hypotheses

Path	Description
logs/base/*.eval	Inspect eval logs for base model
logs/sft/*.eval	Inspect eval logs for SFT model
logs/custom_sft/*.eval	Inspect eval logs for custom SFT model
hypotheses/*.jsonl	Generated behavior hypotheses

Extraction details

• Layer index: 16 (midlayer)
• Hidden dim: 4096
• Max sequence length: 8192
• Precision: float32 (extracted from bfloat16 model)

Usage

from datasets import load_dataset

# Training doc activations (example for base + custom_sft)
base_acts = load_dataset("jrosseruk/dare-activations", data_files="base/activations.parquet", split="train")
csft_acts = load_dataset("jrosseruk/dare-activations", data_files="custom_sft/activations.parquet", split="train")

# Query activations
base_q = load_dataset("jrosseruk/dare-activations", data_files="base/query_activations.parquet", split="train")
csft_q = load_dataset("jrosseruk/dare-activations", data_files="custom_sft/query_activations.parquet", split="train")