jucamohedano's picture
docs: add dataset card for results archive
3de282d verified
|
Raw History Blame Contribute Delete
5.69 kB
metadata
license: mit
language:
  - en
tags:
  - analysis-results
  - representational-similarity-analysis
  - rsa
  - fmri
  - expert-neurons
  - gpt-2
  - mitchell-2008
  - interpretability
  - brain-alignment
  - chain-of-thought

experts-ml-selfcond-result

This dataset contains the complete results archive for the experts-ml-selfcond research project — a fork of Apple's selfcond (ICML 2022) extended with fMRI brain alignment experiments via Representational Similarity Analysis (RSA).

⚠️ This is a companion results artifact, not a training dataset. It contains analysis outputs (RDMs, RSA statistics, expert neuron tables, visualizations) generated from the pipeline. See jucamohedano/Qwen3-30B-A3B-Instruct-2507_custom_60_cot for the input Chain-of-Thought sentences.

What's Inside

The archive (results_complete.tar.xz) contains the full results/ directory (8.6 GB uncompressed, 1.6 GB compressed) with 55,282 files spanning all pipeline stages:

Directory Structure

results/
├── brain_rdm/              # Brain RDMs per subject × grid region
│   └── .../default/        # Original (non-normalized) brain RDMs
│   └── .../voxel_normalized/     # Per-voxel z-scored (normalization robustness check)
│   └── .../voxel_normalized_original_grid/  # Normalized, original grid shape
├── compute_responses/      # Cached model activations (GPT-2, 12 layers)
├── compute_expertise/      # AP scores per neuron per concept (expertise.csv, .json)
├── unique_experts/         # Filtered unique expert neuron sets (correlation thresholded)
├── expert_overlap/         # Jaccard similarity between concept expert sets
├── word_features/          # Per-layer feature matrices [60 words × |E_l| neurons]
├── subspace_gaze/          # PCA/LAT concept subspaces & UMAP visualizations
├── steering/               # Steering vector outputs & generation samples
├── steering_validation/    # Causal validation of expert vectors
├── rsa/                    # RSA comparison results
│   ├── .../default/        # Original RSA: expert vs. dense embeddings
│   └── .../voxel_normalized_*/  # Voxel-normalized reproducibility runs
├── *.png                   # Aggregate visualizations (heatmaps, scatter plots, etc.)
└── *.json                  # Hydra configs, metadata, run logs

Key Result Types

File Description
brain_rdms.pkl Per-subject brain RDMs (60×60, 1−Pearson r)
expertise.csv AP scores, forcing values per neuron per layer
rsa_results.csv RSA statistics: Spearman rho, Cohen's d, p-values, FDR correction
*_rdm.npy Model RDMs (60×60) per layer, per condition (expert / dense)
*_features.npy Word feature matrices [60 words × N experts]
*.png Heatmaps, effect size plots, scatter comparisons

Key Scientific Results

Original Analysis (non-normalized brain RDMs)

  • MLP projection layer (mlp.c_proj) at AP ≥ 0.6: expert neurons significantly outperform dense embeddings (mean Cohen's d = +0.627)
  • Attention layers require stricter filtering; MLP layers align better than attention

Voxel Normalization Robustness Check

Per-voxel z-scoring was applied as a methodological control (see scripts/compute_brain_rdm.py):

Layer Original Mean d Voxel-Normalized Mean d
mlp.c_proj (AP=0.6) +0.627 −0.523
attn.c_proj (AP=0.7) +0.536 −0.551
mlp.c_fc (AP=0.6) +0.443 −0.010
attn.c_attn (AP=0.7) +0.412 −0.522

Finding: Per-voxel normalization reverses or eliminates the expert neuron advantage across all layer components, suggesting high-variance voxels were driving the original brain alignment signal. See my_docs/voxel_normalization_analysis.md for the full analysis.

How to Use

pip install huggingface_hub

# Download the tarball (requires access to this private dataset)
from huggingface_hub import hf_hub_download

hf_hub_download(
    repo_id="jucamohedano/experts-ml-selfcond-result",
    filename="results_complete.tar.xz",
)

# Extract
tar -xJf results_complete.tar.xz

Alternatively, browse individual files on the Hugging Face dataset page.

Reproducing the Analysis

The pipeline is orchestrated via Hydra in run_pipeline.py:

python run_pipeline.py task=brain_rdm            # Compute brain RDMs
python run_pipeline.py task=compute_responses     # Cache model activations
python run_pipeline.py task=compute_expertise     # Compute AP scores
python run_pipeline.py task=word_feature_extraction  # Extract expert/dense features
python run_pipeline.py task=rsa                   # RSA comparison

Citation

If you use the original self-conditioning methods, cite:

Xavier Suau, Luca Zappella, Nicholas Apostoloff. Self-Conditioning Pre-Trained Language Models. ICML 2022.

If you use the brain-alignment/RSA extensions, refer to this repository and the final report: JuanCamachoMohedano-257536-Project-Experts.pdf.

Related Datasets