Commit History
-
Major update: Add MLP→CLT mappings + retrained transcoders
KokosDev- Added mapping_L{0-30}.pt files with MLP→CLT correlation mappings - Retrained all 31 transcoders with improved training (5000 steps each) - Updated README with comprehensive usage documentation - Added training_summary.json with detailed metrics - All layers achieve 0% dead features - Sparsity: 2-29% L0 (layer-dependent, deeper layers sparser) - Reconstruction loss: <0.1 for most layers Key features: ✅ MLP→CLT co-activation based mapping for feature attribution ✅ Decoder weights for CLT→MLP reconstruction ✅ Complete training metadata and metrics ✅ Ready for circuit-tracer style attribution analysisbc702e3