GitHub - Cacapice/TransformerLens-Training-Controls: TransformerLens compatibility layer for a maritime routing model trained from scratch, with numerically verified weight conversion preserving logits exactly. Enables activation patching, residual-stream analysis, and linear probing to study whether adversarial routing intent is encoded before execution.

Branches Tags

Name		Name	Last commit message	Last commit date
Latest commit History 2 Commits
PR_BODY.md		PR_BODY.md
__init__.py		__init__.py
_pretrain_mocks.py		_pretrain_mocks.py
architecture_adapter_factory.py		architecture_adapter_factory.py
bridge.py		bridge.py
moe.py		moe.py
pretrain.py		pretrain.py
test_bridge_train_mode_propagation.py		test_bridge_train_mode_propagation.py
test_model_registry.py		test_model_registry.py
test_moe_bridge_tuple_output.py		test_moe_bridge_tuple_output.py
test_pretrain_adapter.py		test_pretrain_adapter.py
test_pretrain_adapter_integration.py		test_pretrain_adapter_integration.py
test_pretrain_model_container.py		test_pretrain_model_container.py
tiny_pretrain_model.py		tiny_pretrain_model.py

About

TransformerLens compatibility layer for a maritime routing model trained from scratch, with numerically verified weight conversion preserving logits exactly. Enables activation patching, residual-stream analysis, and linear probing to study whether adversarial routing intent is encoded before execution.