Closed dtch1997 closed 2 months ago
[!WARNING]
Review failed
The pull request is closed.
The project saw the introduction of a new example notebook and various enhancements to the existing codebase. Notebook additions provide a comprehensive pipeline for model handling, graph operations, and ablation experiments. The code changes encompass renaming and refining functions and classes, revising method signatures, and introducing new utilities for cache management and graph pruning. These updates augment functionality and usability and improve cohesion between modules.
Files/Groups | Summary |
---|---|
.gitignore |
Added *.pkl to ignore list. |
examples/walkthrough.ipynb |
Introduced a pipeline for model loading, graph operations, and ablation experiments. |
sae_eap/attribute.py |
Renamed several classes and functions for clarity. |
sae_eap/cache.py |
Created a CacheHook namedtuple, added new cache-management functions, and renamed existing ones. |
sae_eap/prune.py |
Updated imports and generalized types in method signatures. |
sae_eap/graph/viz.py |
Added new imports and modified edge visualization logic. |
sae_eap/runner.py |
Replaced manual cache hook creation with a utility function. |
tests/unit/attribute/test_attribute.py |
Introduced test function for model cache comparison. |
tests/unit/test_prune.py |
Added test cases for edge sorting and pruning in a graph. |
tests/integration/attribute/test_attribute.py |
Removed one test and added another to validate forward hooks. |
tests/unit/attribute/test_cache_hooks_and_dicts.py |
Added tests for cache hooks and dictionaries in model transformations. |
sequenceDiagram
participant User
participant Notebook as examples/walkthrough.ipynb
participant ModelLoader as sae_eap/model/load_pretrained
participant Graph as sae_eap/graph/build
participant Attribution as sae_eap/runner
participant Pruner as sae_eap/prune
participant Viz as sae_eap/graph/viz
User ->> Notebook: Open notebook
Notebook ->> ModelLoader: Call load_model
ModelLoader -->> Notebook: Return model
Notebook ->> ModelLoader: Call make_ioi_single
Notebook ->> Graph: Call build_graph
Graph -->> Notebook: Return graph
Notebook ->> Attribution: Call run_attribution
Notebook ->> Pruner: Call PruningPipeline
Pruner -->> Notebook: Return pruned graph
Notebook ->> Viz: Call to_graphviz
Viz -->> Notebook: Return visualized graph
Notebook ->> Attribution: Call run_ablation
Notebook ->> Viz: Call plot_faithfulness_curves
In the world of code, vast and grand,
A notebook's journey, a careful hand,
Model and graphs in harmony dance,
Cache hooks and pruners at first glance.
Tests now guard our precious scripts,
Ensuring no bugs sneak through the crypts.
Hail to progress, in bytes we trust,
Onward, advancements, robust and just.
[!TIP]
Early access features: enabled
We are currently testing the following features in early access: - **OpenAI `gpt-4o` model for code reviews and chat**: OpenAI claims that this model is better at understanding and generating code than the previous models. We seek your feedback over the next few weeks before making it generally available. Note: - You can enable or disable early access features from the CodeRabbit UI or by updating the CodeRabbit configuration file. - Please join our [Discord Community](https://discord.com/invite/GsXnASn26c) to provide feedback and report issues. - OSS projects are currently opted into early access features by default.
Thank you for using CodeRabbit. We offer it for free to the OSS community and would appreciate your support in helping us grow. If you find it useful, would you consider giving us a shout-out on your favorite social media?
Summary by CodeRabbit
New Features
Bug Fixes
Tests
Chores