v2.0.0

What's Changed

• Edge Ablations by @nix-apollo in #285
• Tweak edge ablation's handling of constant directions by @nix-apollo in #297
• Support jacobian basis (formerly November B): Refactor version. by @stefan-apollo in #294
• resupport tinystories by @nix-apollo in #304
• Add helper function for interaction rotation operation by @nix-apollo in #306
• [RFR] Bisect ablation scheduler by @stefan-apollo in #303
• Format with new black and pin dev dependencies by @danbraunai-apollo in #309
• Run ablation tests without reading sample graphs by @danbraunai-apollo in #307
• Split layernorm into two sequential modules by @danbraunai-apollo in #305
• Small plotting fixes by @nix-apollo in #308
• Stochastic sources jacobian basis by @stefan-apollo in #296
• Add n_documents option for HF datasets by @danbraunai-apollo in #316
• Gauss-Legendre alpha integral, plus integral refactor by @stefan-apollo in #317
• Allow distributed edge splitting over out_dim by @danbraunai-apollo in #319
• Specify integral method by layer by @nix-apollo in #325
• Build test fixes by @nix-apollo in #330
• Updates for python 3.12 by @nix-apollo in #327
• Distributed calculation for basis by @nix-apollo in #332
• Nix' magic memory saving changes by @stefan-apollo in #338
• Bugfix for distributed seeds by @nix-apollo in #340
• Fix/negative variance edge ablation [Replacement for #323] by @stefan-apollo in #346
• Naive implementation of gradient flow (in refactored code) by @stefan-apollo in #333
• Isolate the variance in layernorm into a separate RIB dir by @nix-apollo in #345
• Feature/plotting improvements by @nix-apollo in #348
• Ablations with reference to baseline by @stefan-apollo in #347
• Fix for plotting bug introduced in #348 by @stefan-apollo in #351
• Use a huggingface cache by @stefan-apollo in #350
• Use separate (larger) dataset for gram (and mean) matrices by @stefan-apollo in #344
• Set huggingface cache at the right point by @stefan-apollo in #356
• Take cache mode activations off the GPU by @stefan-apollo in #355
• Improve ablations by @nix-apollo in #357
• Improved docstring / comments by @stefan-apollo in #359
• Update the modular addition config file by @stefan-apollo in #360
• Added new hook & function for feature viz tool by @stefan-apollo in #363
• Preparation for publication by @danbraunai-apollo in #364
• Modularity Analysis by @nix-apollo in #349
• Paper plotting scripts by @stefan-apollo in #366
• Implement paper plots feedback by @stefan-apollo in #367

Full Changelog: v1.0...v2.0.0

v1.0

Breaking Changes

Many. The most important being that:

1. All rib build experiments were combined and can be run fromrib_scripts/rib_build and all ablation experiments were combined and can be run from rib_scripts/ablations.
2. Rib build outputs a RibBuildResults pydantic object.

Some more detailed breaking changes are given in the PR descriptions that were merged for this release.

What's Changed

• Fix bug overwriting edges each batch by @danbraunai-apollo in #243
• Separated out attention scores into module to be cache-able by @stefan-apollo in #246
• Add option to add node_labels to graph by @stefan-apollo in #250
• Normalize after squaring in jacobian squared by @danbraunai-apollo in #247
• Throw a warning if only the IGNORE field is mismatched by @stefan-apollo in #252
• Utility to get rib activations by @nix-apollo in #242
• Update workflow file for self-hosted runner by @danbraunai-apollo in #255
• Implement ModularDNN model by @stefan-apollo in #235
• Add unused import error in pre-commit by @danbraunai-apollo in #258
• Add get_config functions to test_build_graph by @nix-apollo in #256
• PCA basis by @nix-apollo in #249
• Remove 'model' key when parsing interaction_graph_info by @danbraunai-apollo in #262
• Combine lm and mlp ablation scripts by @danbraunai-apollo in #263
• Combine lm_rib_build with mlp_rib_build by @danbraunai-apollo in #264
• Move core functionality from experiments to rib package by @danbraunai-apollo in #267
• Make RibResults a pydantic model by @danbraunai-apollo in #268
• Use raw MLP class for modular MLPs by @danbraunai-apollo in #269
• Convert InteractionRotation and Eigenvectors to pydantic models by @danbraunai-apollo in #270
• Store edges in Edges pydantic model by @danbraunai-apollo in #273
• Rename variables and clean docs by @danbraunai-apollo in #274
• Add Stochastic Noise Sources (SNS) attribution method by @danbraunai-apollo in #254
• add kwarg support for test configs by @nix-apollo in #277
• Move ablation schedule from fn to instance methods by @danbraunai-apollo in #281
• Nix's unify clenaups by @nix-apollo in #280
• Centered rib by @nix-apollo in #257
• Rename torch types by @danbraunai-apollo in #284
• Remove extra return_set arg by @danbraunai-apollo in #286
• Combine Eigenvectors object with InteractionRotation by @danbraunai-apollo in #290
• Major Refactoring by @danbraunai-apollo in #260

Full Changelog: v0.2...v1.0

v0.2

What's Changed

• Support CIFAR model training by @nix-apollo in #223. Breaking Changes:
  • Moved and renamed files and folders that referenced MNIST
  • Moved the mlp experiment file's ModelConfig, that is now MlpConfig in models/mlp.py
  • MLP configs require input and output size.
• Support different bases and edges formulas by @stefan-apollo in #219
• Forbid extra args in pydantic configs by @danbraunai-apollo in #237. Breaking Changes
  • Any configs that contained arguments which weren't valid for that config will not throw a pydantic ValidationError.
• Tiny stories support by @nix-apollo in #232
• Add svd basis by @nix-apollo in #239
• Support return_set_frac and n_samples for all datasets by @danbraunai-apollo in #238. Breaking Changes
  • No longer supports passing "both" as a return_set argument.
  • Previously, VisionDataset had a return_set_frac that took samples from the start of the dataset. This change will therefore be a breaking change for configs that previously defined a return_set_frac. VisionDataset is a new thing, and Nix said that this isn't a concern.

Full Changelog: v0.1...v0.2

v0.1

Breaking Changes

Update transformer-lens (#222)

• You must re-install the rib package in your environment as the requirements have changed.

MLP complete rewrite in #201

• MLP layers and models have an updated interface and fold biases in a very different way. In particular, note:
  • You can't train after folding the bias
  • You can't fold the bias of sigmoid or tanh models, as there is no x where act(x)=1 for these functions.
  • MLP models and layers now have a .has_bias_folded and .fold_bias attribute. This is the preferred interface for bias folding.
  • Renamed mlp Layer class to MLPLayer (especially as it's imported into rib.models namespace)
• Old MLP models will no longer be able to be loaded.

Return values and out_dirs (#213)

• Previously a main method would return None if the user decided not to overwrite the file. Now it raises a FileExistsError. This is so the main method can have a non-optional return type.
• changed the meaning of out_dir=None to not write an output file. The default is still .out/
• AblationConfigs can no longer have no experiment name. Previously this was used to have no output file. In able to have no output file, pass out_dir = None instead.

Can no longer pass eps to an lm_rib_build or lm_ablation config (#225)

What's Changed

• Parallelize edge calculations by @nix-apollo in #196
• Fold biases by default by @danbraunai-apollo in #199
• main methods accept config objects in addition to paths by @nix-apollo in #200
• Feature to force overwrite of output file via cmdline argument or config file by @stefan-apollo in #202
• MLP class refactor, particularly for fold_bias by @nix-apollo in #201
• Fix M matrix floating point issues by @stefan-apollo in #203
• Assert sequential node_layers in run_mnist_rib_build by @danbraunai-apollo in #214
• Clean up tests and standardize configs by @nix-apollo in #213
• Run RIB build over multiple nodes by @danbraunai-apollo in #208
• Clean up dtype loading by @danbraunai-apollo in #215
• Update tlens and misc config settings by @danbraunai-apollo in #222
• Add n_ctx to config by @nix-apollo in #209
• Remove epsilon from script configs by @danbraunai-apollo in #225

New Contributors

• @nix-apollo made their first contribution in #196

Full Changelog: v0.0...v1.0

v0.0

State of Project

• Runs single-process RIB on MNIST, modular addition, and Pythia-14m (although the latter will take many days/weeks to compute all edges).
• Calculates ablation curves

Recent Updates

Normalize gradients in edge calculation #189
Normalizes gradients by the dataset_size and token dimension before summing with the input tensor

Implement split-up Attention module #180
Splits Atttention (attn) into AttentionIn (attn_in) and AttentionOut (attn_out). attn_in just creates the KQV activations, and attn_out handles the rest

Replace jacrev with for loop #185
Uses a for loop instead of jacrev for the edge calculation

Refactor SequentialTransformer for cleanliness #183
Refactors SequentialTransformer and major cleanup of it's docs and structure. I recommend having a read of the full file, it might clear things up for you about how this class works and what module_id and section_ids are.

Map for module_id to section_id mapping #182

Introduces a mapping function between a section_id (e.g. sections.section_2.3 to a module_id (e.g. mlp_out.3 ). E.g. you can do:

>> model = SequentialTransformer(cfg, node_layers)
>> print(model.section_id_to_module_id["sections.section_2.3"])
mlp_out.3