File size: 407 Bytes
cb5d220 2da9072 cb5d220 | 1 2 3 4 5 6 7 8 | # Atomight-V1-GRAFT-0.8M
Custom from-scratch hybrid discriminative+generative model for abstract reasoning (RAVEN / PGM / I-RAVEN-X style matrix puzzles).
Architecture: GRAFT (Grid Relational Attention Fusion Transformer), ~0.81M params. FiLM-conditioned narrator, anti-collapse variance regularization on the rule embedding.
Load with the GRAFT class definition (see modeling.py) + model.pt state_dict.
|