SAE Labeling

• AutoInterp Lite (Feature Discovery): Quickly rank domain-relevant features out of thousands (minutes).

• AutoInterp Full (Feature Explanation): Produce human-readable explanations with confidence & detection metrics for specific features or all features (deeper, slower).

1. Lite → find top 10–50 candidates for your domain.

2. Full → generate vetted explanations and scores for those feature IDs (or sweep all).

• Trained SAE (weights + config)

• Base model used for activations (e.g., bert-base-uncased, meta-llama/...)

• Two corpora: domain (e.g., financial news) and general (e.g., Wikipedia)

• Ranked CSV of features with:
• Specialization score (domain vs general)
• Activation strength (mean activation)
• Top activating examples (text spans)
• Optional proto-label (heuristic string from examples)

• Specialization score ↑ = more domain-specific

• Domain activation vs General activation

• Specialization confidence (stability of the rank under resampling)

• 1k–10k features in 2–5 minutes on a single GPU (sampling-based)

bash
cd autointerp_lite
python run_analysis.py \
  --sae_path /models/sae/layer6 \
  --base_model bert-base-uncased \
  --domain_corpus data/finance.jsonl \
  --general_corpus data/general.jsonl \
  --out_dir results/lite_finance

Rule of thumb: specialization > 3.0 and confidence > 30 are strong candidates.

• Uses a chat-formatted LLM to generate explanations from activating vs non-activating examples.

• Adds contrastive analysis (FAISS) so the LLM sees near misses and avoids vague labels.

• Computes detection metrics (F1 / Precision / Recall) for quality assurance.

• Feature IDs from Lite (e.g., 27 133 220) or “analyze all”

• Explainer chat model (cloud or local)

• Optional FAISS index for contrastive negatives

• Explanation cards (Markdown/JSON) per feature

• Scores: F1, Precision, Recall (+ per-feature thresholds)

• Examples: positive/negative spans used for validation

• 30–60 min/feature (depends on dataset size & LLM)

bash
cd autointerp_full
./example_LLM_API.sh \
  --sae_path /models/sae/layer6 \
  --base_model bert-base-uncased \
  --feature_num 27 133 220 \
  --n_tokens 200000 \
  --explainer_model openai/gpt-4o-mini \
  --non_activating_source FAISS \
  --out_dir results/full_finance

plain text
results/full_finance/
  explanations/FEATURE_133.md
  scores/detection/FEATURE_133.json  # f1, precision, recall, threshold
  results_summary.csv

• F1 ≥ 0.70, Precision ≥ 0.80, Recall ≥ 0.60 → production-worthy

1. Embed text spans (sentence-transformers)

2. Build FAISS index over non-activating spans

3. For each feature, retrieve closest negatives to pair with positives

4. Prompt LLM with contrastive pairs → sharper, less vague explanations

bash
# From repo root
pip install -e .

• SAE model path (trained previously)

• Base model path/name (for activation extraction)

• Domain & general corpora (for Lite)

• Chat model access (API key or local checkpoint) for Full

bash
cd autointerp_lite
python run_analysis.py \
  --sae_path /models/sae/layer6 \
  --base_model bert-base-uncased \
  --domain_corpus data/finance.jsonl \
  --general_corpus data/general.jsonl \
  --out_dir results/lite_finance

bash
cd autointerp_full
./example_LLM_API.sh \
  --sae_path /models/sae/layer6 \
  --base_model bert-base-uncased \
  --feature_num 27 133 220 \
  --explainer_model openai/gpt-4o-mini \
  --non_activating_source FAISS \
  --out_dir results/full_finance

plain text
feature_id:int, proto_label:str, specialization:float,
domain_activation:float, general_activation:float, specialization_conf:float,
top_examples:list[str]

json
{
  "feature_id": 133,
  "label": "Earnings reports / rate changes",
  "f1": 0.74,
  "precision": 0.82,
  "recall": 0.67,
  "threshold": 0.43,
  "positive_examples": ["..."],
  "negative_examples": ["..."]
}

• Use Lite when you have thousands of features and need a shortlist (10–50) for your domain.

• Use Full when you need validated semantics with metrics and contrastive evidence (for research, audits, safety).

• Balance evidence size: start with -n_tokens 2e4 to iterate quickly; scale up later.

• Contrastive negatives: prefer FAISS over random; explanations are crisper.

• De-duplication: similar features may exist; use cosine on SAE atoms to merge siblings.

• Guardrails: exclude features with F1 < 0.5 from production catalogs.

• Reproducibility: fix seeds and persist corpora snapshot for consistent labels.

• AutoInterp Lite README – feature ranking & discovery

• AutoInterp Full README – LLM-based explanation, contrastive prompting, scoring