HUSAI: Reliability-First SAE Research

HUSAI studies whether sparse autoencoder (SAE) features are trustworthy under strict release criteria: internal reproducibility, stress robustness, and external benchmark competitiveness.

Key Finding

Trained SAE features are indistinguishable from random baseline (PWMCC = 0.309 vs 0.300 for untrained SAEs). SAEs reconstruct well but learn arbitrary, non-reproducible feature decompositions across seeds. See paper/sae_stability_paper.md for the full paper.

Current Bottom Line

• Internal consistency signal: positive.
• Stress controls (random_model, transcoder, OOD): passing.
• External transfer (SAEBench, CE-Bench): below strict thresholds.
• Strict release outcome: pass_all=false.

Use EVIDENCE_STATUS.md before citing exact metrics (local vs remote evidence tiers).

Start Here

See START_HERE.md for the full reading order and orientation.

Quick Validation

conda env create -f environment.yml && conda activate husai
pip install -r requirements-dev.txt
pytest tests -q
make smoke

Core Scripts

Internal baselines and ablations:

• scripts/experiments/run_phase4a_reproduction.py
• scripts/experiments/run_core_ablations.py
• scripts/experiments/run_assignment_consistency_v3.py

Follow-up experiments (Section 4.11 of the paper):

• scripts/experiments/run_all_followup_experiments.sh -- run all follow-ups
• scripts/experiments/exp_1layer_ground_truth.py -- 1-layer vs 2-layer comparison
• scripts/experiments/exp_subspace_stability.py -- subspace vs feature stability
• scripts/experiments/exp_effective_rank_predictor.py -- universal stability predictor
• scripts/experiments/exp_contrastive_stability.py -- contrastive alignment loss
• scripts/experiments/exp_intervention_stability.py -- steering consistency across seeds
• scripts/experiments/exp_dictionary_pinning.py -- warm-start with frozen decoder columns
• scripts/experiments/exp_pythia70m_stability.py -- scale to Pythia-70M (GPU)

External benchmark program:

• scripts/experiments/run_husai_saebench_custom_eval.py
• scripts/experiments/run_husai_cebench_custom_eval.py
• scripts/experiments/run_architecture_frontier_external.py

Strict gating:

• scripts/experiments/select_release_candidate.py
• scripts/experiments/run_stress_gated_release_policy.py

License

MIT (LICENSE).