SetConCA Generalist

Trained SetConCA sparse dictionary checkpoint for the paper.

Expected files

model.safetensors or setconca_latest.pt
config.json
metrics.json
eval_table_fullbank.csv
eval_table_geometry.csv
eval_table_lovo.csv

Configuration

  • input activation dimension: 2048
  • encoder dimension: 1024
  • concept dimension: 8192
  • TopK: 256
  • aggregation: precision-weighted Gaussian
  • objective: L_set_rec + 1.5 * L_p2s
  • temperature: 0.10

Intended use

Research on sparse representation learning, grouped activation-set geometry, and mechanistic interpretability.

Limitations

The learned sparse coordinates are not guaranteed to be monosemantic and are not validated as causal intervention targets.

Downloads last month
6
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support