Easy Sparse Linear Autoencoder (https://arxiv.org/abs/2309.08600) training, with data generated from TransformerLens models. This code represents a simplification of the codebase actually used in the paper, and is designed significantly better and is much more readable. The legacy code is available in the legacy branch.

Hopefully all the neccesary packages should be in requirements.txt.

To sample activations, run generate_test_data.py with the flags

• --model [str] to specify which model to use (using TransformerLens naming)
• --n_chunks [int] to specify how many chunks (files containing activations) to generate
• --chunk_size [int] to specify the size of the chunks in activations
• --dataset [str] to specify the HuggingFace dataset to run on
• --locations [str] to specify which activation to sample, using TransformerLens hook naming.
• --dataset_folder [str] to specify the output folder
• --device [str] to specify a PyTorch device to run the model on

Some of these flags have useful defaults, which you can see in the python file.

To train an autoencoder, run basic_l1_sweep.py with the flags

• --dataset_dir [str] to specify the activation dataset
• --output_dir [str] to specify where to save the models to
• --ratio [float] to specify the 'blowup factor' (features / activation dimensions)
• --l1_value_min [float] to specify the minimum L1 penalty factor (log10)
• --l1_value_max [float] to specify the maximum L1 penalty factor (log10)
• --batch_size [int] to specify the training batch size
• --device to specify the PyTorch device to train on
• --adam_lr to specify the Adam learning rate
• --n_repetitions to specify how many times to train on the dataset
• --save_after_every to toggle from saving after every repetion (including the first) to saving after every chunk

Again, some of these flags have useful defaults.

Trained SAEs are outputted as instances of the SparseLinearAutoencoder class (defined in training/dictionary.py), as a dictionary indexed by l1_penalty.

Internally, the sweeps over L1 penalty ranges are implemented using a model ensembler defined in training/ensemble.py. It should be robust to most modifications of autoencoder architecture, but you might have to fiddle with it if you make strange changes.

python generate_test_data.py --model="EleutherAI/pythia-70m-deduped" --layers 2 --n_chunks=2
python basic_l1_sweep.py --dataset_dir="activation_data/layer_2" --output_dir="output_basic_test" --ratio=8 --batch_size=4096 --n_repetitions=2 --save_after_every