transformerlensorg/transformerlens is a mechanistic interpretability library that dissects the computations GPT-family language models acquire during training. It is written in Python and released under the MIT license. The latest release as listed by the repository is v4.2.0, and the changes contained in this version are not given in the available information.
The goal stated in the README is to analyze the weights of a trained model and reverse-engineer the algorithm it learned during training. Internal activations are intermediate values computed as the input passes through the layers; by loading the model with TransformerBridge, these can be inspected and cached. Running activations can be edited, removed, or replaced, so you can observe and fix what computation takes place inside the model. By default it keeps HuggingFace's raw weights. The repository was created on August 26, 2022 and has 3,983 stars.
However, the published information contains no performance figures, no benchmark results, and no release notes for v4.2.0. The research database records only a momentum score of 3.0, and the scores for the other three items are empty. The only thing confirmed by the available information is that it is a tool for observing and editing the internal computation of trained models. How activation editing affects actual output is difficult to judge from this information alone.