ApolloResearch/rib

Issues

• Error in distributed test

  #365 opened 7 months ago by danbraunai-apollo
  0

• Set tensors to cpu in _append_to_hooked_list

  #278 opened 8 months ago by danbraunai-apollo
  0

• `create_hf_dataset` optional argument `model_n_ctx` is not optional

  #354 opened 8 months ago by stefan-apollo
  0

• naive_gradient_flow and basis_formula=svd should confict

  #352 opened 9 months ago by stefan-apollo
  0

• Node spacing broken for 1st layer?

  #271 opened 9 months ago by stefan-apollo
  1

• Some tests fail on A100 (that do not fail on A6000)

  #339 opened 9 months ago by stefan-apollo
  0

• Existing outfiles (check_outfile_overwrite = False) is handled inconsistently

  #336 opened 9 months ago by stefan-apollo
  0

• load_interaction_rotations should return calc_Cs_time

  #335 opened 9 months ago by stefan-apollo
  0

• Fix flaky test `test_pythia_14m_build_graph_jacobian_stochastic`

  #329 opened 9 months ago by nix-apollo
  0

• Investigate rotate_final_layer_invariance test

  #287 opened 9 months ago by danbraunai-apollo
  0

• Negative variance iff edge ablations in Split LN

  #318 opened 9 months ago by stefan-apollo
  0

• Make RIB more memory efficient

  #324 opened 9 months ago by nix-apollo
  0

• Cache transformers datasets

  #320 opened 9 months ago by danbraunai-apollo
  1

• Update yaml files in repo

  #283 opened 10 months ago by stefan-apollo
  1

• Fix VSCode black and venv mismatch

  #311 opened 10 months ago by danbraunai-apollo
  1

• Implement token masking to exclude BOS token

  #313 opened 9 months ago by LuciusApollo
  1

• Test RIB on the test set rather than train set

  #314 opened 9 months ago by stefan-apollo
  0

• Stray edges in squared edge attribution / numerical issues with large Lambda[0] values

  #276 opened 10 months ago by stefan-apollo
  2

• Better way to specify node layers

  #312 opened 9 months ago by nix-apollo
  0

• Split layernorm into two layers

  #299 opened 10 months ago by danbraunai-apollo
  0

• Simplify test_ablation setup

  #291 opened 10 months ago by danbraunai-apollo
  1

• Update our default kwarg values

  #310 opened 10 months ago by stefan-apollo
  0

• Make load_dataset instance methods of dataset configs

  #289 opened 10 months ago by danbraunai-apollo
  1

• Make C calculation more memory efficient

  #300 opened 10 months ago by danbraunai-apollo
  1

• Add option to start integrals from the expected value

  #301 opened 10 months ago by nix-apollo
  0

• get_rib_acts is batch-size dependent (floating point error)

  #295 opened 10 months ago by stefan-apollo
  0

• Bug in orthogonal ablations with centered rib

  #292 opened 10 months ago by nix-apollo
  0

• Edge Ablations

  #279 opened 10 months ago by nix-apollo
  0

• Read nodes_per_layer from file in plot_mlp_graph

  #236 opened 10 months ago by stefan-apollo
  4

• Make RibResults a pydantic model

  #261 opened 10 months ago by danbraunai-apollo
  0

• Put rib_build experiment inside rib package

  #265 opened 10 months ago by danbraunai-apollo
  0

• Explicitly validate node layers for mlp builds

  #266 opened 10 months ago by danbraunai-apollo
  0

• Combine rib build and ablation scripts for lms and mlps

  #228 opened 10 months ago by danbraunai-apollo
  1

• Ablations Schedules should return num_to_keep not num_to_ablate

  #288 opened 10 months ago by nix-apollo
  0

• Support rib centering

  #248 opened 10 months ago by nix-apollo
  0

• Return dictionaries of custom objects rather than lists

  #282 opened 10 months ago by danbraunai-apollo
  0

• Create Sequential Component ABC with module properties

  #231 opened 10 months ago by danbraunai-apollo
  4

• PCA basis beats RIB in some Mod Add layers

  #275 opened 10 months ago by stefan-apollo
  2

• MLP rib build -> ablation pipeline is broken.

  #259 opened a year ago by nix-apollo
  3

• `load_sequential_transformer` device argument doesn't work

  #220 opened a year ago by nix-apollo
  2

• Pythia attention_scores mismatch (1e-5) when folding biases

  #245 opened a year ago by stefan-apollo
  0

• Capture warnings in logger

  #253 opened a year ago by nix-apollo
  2

• Prevent assignment to all configs

  #240 opened a year ago by danbraunai-apollo
  2

• Add return_set_frac and return_set_n_samples to modadd

  #234 opened a year ago by danbraunai-apollo
  0

• Mod add seed issue

  #229 opened a year ago by nix-apollo
  2

• Graphs not invariant under `rotate_final_node_layer`

  #241 opened a year ago by stefan-apollo
  0

• Prevent extra arguments to Dataset classes

  #233 opened a year ago by danbraunai-apollo
  1

• Remove eps parameter in the config

  #224 opened a year ago by danbraunai-apollo
  0

• Replace "scaler" and "interval_size" by general weights

  #226 opened a year ago by stefan-apollo
  0

• Train, ribify and analyse modadd with layernorm

  #221 opened a year ago by danbraunai-apollo
  1