joshycodes/activation-control-battery

activation-control battery: stored residual-stream activations

Raw recordings from the activation-control battery: for each model, residual-stream activations over the transcribed sentence span for every trial of the instruction battery (think about X / don't think about X / intensity k of 4 / no instruction, 10 concepts, ~8600 trials per model), recorded at 20 depths. Each model has a main run and a layer-targeting (_lt) companion run: the same recording setup with the layer_location condition group (4 conditions) in place of the 19-condition main battery. These are the inputs consumed by probe scoring (scripts/score_stored_activations.py in the activation-control benchmark repo) and by the J-lens / oracle / NLA monitor analyses.

Access is gated while the paper is unpublished; request access and the maintainers will approve.

Layout

raw/<run_name>/config.yaml               battery run config
raw/<run_name>/results.json              per-trial generations and scores
raw/<run_name>/no_instruction_cache.pkl  cached no-instruction baseline (main runs only)
raw/<run_name>/results.pkl               all trials + activations  (files <= 45 GB)
raw/<run_name>/results.pkl.part-NNN      25 GB byte-range parts    (files > 45 GB)
manifest.json                            sizes, part counts, sha256 for split files

Reassemble split files and verify against manifest.json:

cat results.pkl.part-* > results.pkl
shasum -a 256 results.pkl   # must match manifest.json runs[<run_name>].sha256

results.pkl schema

Top-level dict:

• results: list of trial dicts
• activation_codec: "bf16-uint16" (activations are bf16 stored as uint16)

Each trial dict has at least: trial_idx, condition_id (e.g. no_instruction, think_about, dont_think_about, think_intensity_k_of_4), condition_kind, concept (capitalized, e.g. Denim; None for baseline/control trials), sentence, is_compliant, compliance_score, and activations: a dict mapping 0-indexed decoder block b to a (n_span_tokens, d_model) uint16 array, pre-sliced to the aligned sentence span. A handful of trials store the string "__dropped__" in place of an array.

Decode activations:

import numpy as np

def unpack_bf16(u16: np.ndarray) -> np.ndarray:
    return (u16.astype(np.uint32) << 16).view(np.float32)

Layer convention: the pkl keys activations by 0-indexed decoder block b; the equivalent HuggingFace hidden_states index is b + 1. Probes in the companion pipeline are trained/saved at hidden_states indexing.

Downloading one run

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="joshycodes/activation-control-battery",
    repo_type="dataset",
    allow_patterns=["raw/20260715_155552_gemma2_9b_activation_control/*"],
    local_dir="battery",
)

Runs

20 models, one main run and one layer-targeting (_lt) run each; ~1.8 TB total.

Main runs

model_key	run	results.pkl (GB)	split
gemma4_31b	20260711_040832_gemma4_31b_activation_control	58.9	yes
qwen36_27b	20260711_190502_qwen36_27b_activation_control	47.8	yes
qwen_72b	20260712_025112_qwen_72b_activation_control	88.1	yes
llama33_70b	20260712_041546_llama33_70b_activation_control	86.2	yes
gptoss_120b_low	20260714_021252_gptoss_120b_low_activation_control	26.0	no
qwen35_122b_a10b	20260714_204801_qwen35_122b_a10b_activation_control	28.8	no
gemma3_27b	20260715_030627_gemma3_27b_activation_control	53.3	yes
gemma2_9b	20260715_155552_gemma2_9b_activation_control	40.6	no
olmo31_32b	20260718_200202_olmo31_32b_activation_control	45.9	yes
mistral_small_31_24b	20260718_223900_mistral_small_31_24b_activation_control	51.6	yes
olmo3_7b	20260722_000046_olmo3_7b_activation_control	38.5	no
qwen35_4b	20260722_000046_qwen35_4b_activation_control	24.0	no
llama_8b	20260722_003238_llama_8b_activation_control	43.7	no
qwen35_9b	20260722_003238_qwen35_9b_activation_control	38.3	no
gemma4_12b	20260722_010419_gemma4_12b_activation_control	42.2	no
glm47_flash	20260722_024102_glm47_flash_activation_control	18.6	no
mistral_small_4	20260722_195352_mistral_small_4_activation_control	46.6	yes
llama4_scout	20260722_204344_llama4_scout_activation_control	51.6	yes
glm46v	20260722_220417_glm46v_activation_control	49.0	yes
gptoss_20b_low	20260723_232738_gptoss_20b_low_activation_control	25.9	no

Layer-targeting runs

model_key	run	results.pkl (GB)	split
gemma4_31b_lt	20260711_042936_gemma4_31b_activation_control_lt	60.2	yes
qwen36_27b_lt	20260711_193048_qwen36_27b_activation_control_lt	49.0	yes
qwen_72b_lt	20260712_031319_qwen_72b_activation_control_lt	89.9	yes
llama33_70b_lt	20260712_044248_llama33_70b_activation_control_lt	87.0	yes
gptoss_120b_low_lt	20260714_033508_gptoss_120b_low_activation_control_lt	26.5	no
qwen35_122b_a10b_lt	20260714_210650_qwen35_122b_a10b_activation_control_lt	29.5	no
gemma3_27b_lt	20260715_152927_gemma3_27b_activation_control_lt	52.7	yes
gemma2_9b_lt	20260715_161238_gemma2_9b_activation_control_lt	42.6	no
olmo31_32b_lt	20260718_204900_olmo31_32b_activation_control_lt	47.2	yes
mistral_small_31_24b_lt	20260718_230318_mistral_small_31_24b_activation_control_lt	51.5	yes
qwen35_4b_lt	20260722_001112_qwen35_4b_activation_control_lt	24.6	no
olmo3_7b_lt	20260722_002054_olmo3_7b_activation_control_lt	37.9	no
qwen35_9b_lt	20260722_004757_qwen35_9b_activation_control_lt	39.2	no
llama_8b_lt	20260722_005503_llama_8b_activation_control_lt	43.6	no
gemma4_12b_lt	20260722_011710_gemma4_12b_activation_control_lt	43.1	no
glm47_flash_lt	20260722_030626_glm47_flash_activation_control_lt	18.9	no
mistral_small_4_lt	20260722_200946_mistral_small_4_activation_control_lt	48.5	yes
llama4_scout_lt	20260722_210516_llama4_scout_activation_control_lt	52.9	yes
glm46v_lt	20260722_230732_glm46v_activation_control_lt	38.4	no
gptoss_20b_low_lt	20260723_234754_gptoss_20b_low_activation_control_lt	26.5	no

Per-run figures are not mirrored here; they remain on the source volume.

Provenance

Recorded by the activation-control battery (run_experiment.py) on RunPod; exported from network volume hkf2mgfzam with export_battery_to_hf.py. Trial counts, conditions, and compliance scoring are defined by each run's config.yaml.