Rosetta-Activations

Updated: 2026-06-15 02:30 UTC

Contrastive activation extractions for 17 semantic concepts across 46 language models, supporting cross-architecture mechanistic interpretability research.

Companion concept pair corpus: jamesrahenry/Rosetta_Concept_Pairs

Papers: forthcoming

Dataset Structure

Rosetta-Activations/
├── rcp_v1/               # Current extraction line — richest data (N≈2000)
│   └── {Model_Name}/
│       ├── calibration_{concept}.npy           # Peak-layer activations (N=2000)
│       ├── calibration_alllayer_{concept}.npy  # All-layer activations (N=2000)
│       └── calibration_{concept}_meta.json     # Extraction provenance
│
├── paper_n250/           # N=250, C=17 paper corpus — 40 models full analysis; 6 frontier caz-only
│   └── {Model_Name}/
│       ├── calibration_{concept}.npy           # Peak-layer activations (N=250)
│       ├── calibration_alllayer_{concept}.npy  # All-layer activations (N=250)
│       ├── calibration_{concept}_meta.json     # Extraction provenance
│       ├── caz_{concept}.json                  # CAZ analysis
│       ├── gem_{concept}.json                  # GEM analysis
│       ├── ablation_{concept}.json             # Layer-local ablation
│       ├── ablation_gem_{concept}.json         # Handoff vs peak ablation
│       ├── ablation_global_sweep_{concept}.json # Fixed-direction global sweep
│       ├── ablation_random_{concept}.json      # Random-direction null (≤17/model; omitted where concept ablation has zero effect)
│       └── patch_{concept}.json                # Activation patching
│
└── results/              # Cross-model aggregate analyses (e.g. concept_evasion)

Tags (use these, not directory copies)

The paper data is intentionally not the largest-N line. paper_n250/ is the N=250 corpus the papers are built on. New and richer activations (N≈2000) accrue on the rcp_v1/ line, but their derived analysis is not yet computed — don't substitute rcp_v1/ for paper reproduction.

Coverage

• paper_n250/ — N=250, C=17. The 40 standard models are complete: .npy + all 7 JSON analysis families (caz, gem, ablation, ablation_gem, ablation_global_sweep, ablation_random, patch) for every concept (ablation_random ≤17 by design — see tree note). Use for paper reproducibility. 6 large models are caz-only (.npy + caz only), hardware-blocked — see below.
• rcp_v1/ — raw N=2000 activations (.npy + _meta.json) for 40 models. Derived analysis (caz/gem/ablation/global_sweep/random) at N=2000 is not yet computed — it exists only at N=250 in paper_n250/. Backfilling the N=2000 analysis onto rcp_v1/ is planned as GPU allows.
• Several large models (Qwen2.5-32B/72B, Gemma "26B-A4B", Llama-3.1-70B, Falcon-40b) are hardware-blocked pending larger GPUs.

Quick Start

Get the current (richest) activations — N≈2000

pip install huggingface_hub
hf download james-ra-henry/Rosetta-Activations \
    --repo-type dataset --revision current \
    --local-dir ~/rosetta_data/ \
    --include "rcp_v1/*"

Reproduce the published paper results — N=250 (full analysis)

hf download james-ra-henry/Rosetta-Activations \
    --repo-type dataset --revision paper-n250 \
    --local-dir ~/rosetta_data/ \
    --include "paper_n250/*"

Array Format

calibration_{concept}.npy — Peak-layer activations

import numpy as np
# paper_n250 (250 pairs)
acts = np.load("paper_n250/EleutherAI_pythia_6.9b/calibration_agency.npy")
# acts.shape → (500, 4096)
pos = acts[:250]   # agentive
neg = acts[250:]   # non-agentive

# rcp_v1 (2000 pairs)
acts = np.load("rcp_v1/EleutherAI_pythia_6.9b/calibration_agency.npy")
# acts.shape → (4000, 4096)

calibration_alllayer_{concept}.npy — All-layer activations

acts = np.load("paper_n250/EleutherAI_pythia_6.9b/calibration_alllayer_agency.npy")
# acts.shape → (32, 500, 4096)
layer_15 = acts[15]

Probe-overfitting warning (rcp_v1): Peak-layer selection in rcp_v1 was determined by CAZ analysis on the same 2000-pair corpus. Always partition into train/val before fitting probes. The paper_n250 fixed split is the cleanest baseline.

Concepts (17)

Models

Extraction Details

• Pooling: last non-padding token (both splits)
• Pair corpus: jamesrahenry/Rosetta_Concept_Pairs
• Extraction code: rosetta_tools

License

MIT