Tags: ApolloResearch/rib
Tags
Support return_set_frac and n_samples for all datasets (#238) * Support return_set_frac and n_samples for all datasets * Speed up tests * Allow for a None seed in train_test_split * Remove unused import * Subset data after splitting and fix dataset seed * Support random subset for ModAdd and Vision datasets * Add test for get_data_subset * Fix docstring in get_data_subset * Aesthetic changes * Make load_dataset return a single dataset only * Freeze all pydantic configs to prevent assignment * Improve docs of tokenizer_name and return_set * Make seed an Optional argument with default 0 everywhere * Delete dead code * xfail test_modular_arithmetic_rotate_final_layer_invariance * Fix wrong batch size arguments * Disallow mismatched frac_train in ModAdd
Normalize gradients in edge calculation (#189) * Replace jacrev with for loop * Remove out_dim_chunk_size hyperparameter * Prevent user inputting a small or large return_set_frac * Remove deprecated vmap-related dimension checks * Revert testing yaml file for edges_pythia * Fix wrong naming in test function * Fix issue of not being able to load an old config * Normalize by pos_dim and dataset_size in edge calc * Type hint test function