Lccqqqqq/sae-locality-data

SAE Locality Data

Raw experimental artefacts and summary figures for two sparse-autoencoder (SAE) feature-locality experiments across six base language models.

Format note. Data files are PyTorch pickles (.pt). Loading them executes arbitrary code via pickle; only load on a trusted machine and be aware that torch.load(..., weights_only=True) will not work for these files because they contain Python dicts/lists, not tensors.

Top-level layout

ctxlen_xmodel/                  # cross-model "entropy vs context length" experiment
└── <timestamp>/
    ├── run_config.json
    └── <preset>/               # one subdir per model preset
        ├── run.sh              # cluster submission script for this preset
        ├── entropy_vs_context_len_<site>_layer<i>_<ts>.pt
        └── entropy_vs_context_len_<site>_layer<i>_<ts>/   # (legacy/empty in some runs)

<preset>/                        # per-preset "entropy comparison" experiment
└── <timestamp>/
    ├── run_config.json
    ├── bench.json
    ├── entropy_comparison_<site>_layer<i>.pt
    └── entropy_plots_<site>_layer<i>/
        ├── batch_index.json
        └── batch_NNN.png       # per-batch entropy plots

figures/                         # summary figures derived from the above
├── entropy_vs_depth_crossmodel_grid_boxplot.png
├── entropy_vs_depth_crossmodel_grid_violin.png
├── entropy_vs_depth__<preset>.png   # one per preset
└── entropy_plots_resid_out_layer<i>_20260414_053350/  # earlier per-batch plots (pythia-70m)

<preset> is one of pythia-70m, qwen2-0.5b, gpt2-small, llama-3.2-1b, gemma-2-2b, llama-3-8b. <site> is the hookpoint name (resid_post, resid, resid_out, ...) and varies per preset.

File schemas

entropy_comparison_*.pt (per-preset experiment)

{
  "batch_results": [
    {
      "batch_idx": int,
      "start_idx": int,                          # offset into the loader's text stream
      "feature_entropies": {feat_idx: float},    # per-feature entropy in bits
      "token_vector_entropy": float,
      "num_active_features": int,
      "feature_influences": {feat_idx: np.ndarray},  # length-N influence vector per feature
      "feature_activations": {feat_idx: np.ndarray},
      "token_vector_influence": np.ndarray,
    },
    ...                                           # one entry per batch (50 by default)
  ],
  "summary": {"site": str, "preset": str, "timestamp": str, "layer": int, ...},
  "config":  {"preset": str, "threshold": float, "total_features": int, ...},
  "plots_dir": str,                              # absolute path on the machine that produced the run
  "batch_start_indices": [int, ...],
}

entropy_vs_context_len_*.pt (cross-model experiment)

{
  "results_by_context_len": {
    ctx_len: {
      "feature_entropies": {feat_idx: float},
      "token_vector_entropy": float,
      "num_active_features": int,
      ...
    },
    ...                                           # one entry per context length (8, 72, 136, ...)
  },
  "summary": {"preset": str, "site": str, "layer": int, "timestamp": str,
              "max_context_len": int, ...},
  "config":  {"preset": str, "threshold": float, "total_features": int,
              "sae_source": str, ...},
  "plots_dir": str,
}

The top-level run_config.json in each ctxlen_xmodel/<timestamp>/ folder records the global parameters of that cross-model run (presets, per-preset max_context_len/step/char_budget, seed, git commit, host).

Figures

figures/ contains summary plots derived from the raw .pt artefacts above:

• entropy_vs_depth_crossmodel_grid_{boxplot,violin}.png — feature-entropy distribution by layer depth, side-by-side across all six presets.
• entropy_vs_depth__<preset>.png — per-preset depth sweep, one figure per model.
• entropy_plots_resid_out_layer<i>_20260414_053350/ — earlier (2026-04-14) per-batch entropy plots for pythia-70m, kept for reference. The corresponding canonical run in pythia-70m/20260427_105943/ supersedes these but uses the same plotting format.

Caveats

• plots_dir inside each .pt and the host field in run_config.json reflect the originating machine and are not portable.
• The entropy_plots_*/ PNG directories are derived artefacts and can be regenerated from the corresponding .pt.
• Symlinks named latest were used locally to point at the most recent run; they are intentionally not included here. The most recent run is the timestamped subdirectory with the largest <timestamp> value.

Citation / contact

For questions, contact the authors of the originating project.