SAE

1️⃣ Train the SAE model

python dna_training.py

2️⃣ Downstream evaluations with Ridge regression

(Edit the three variables in each script)

A. First-token latent embedding

Edit the following variables in dna_sequence_analysis_firsttoken.py:

csv_path            = "<PATH_TO_YOUR_DATASET_CSV>"
sae_checkpoint_path = "<PATH_TO_YOUR_SAE_CHECKPOINT>"
output_path         = "<PATH_WHERE_YOU_WANT_FIRSTTOKEN_WEIGHTS_CSV>"

Then run:

python dna_sequence_analysis_firsttoken.py

B. Mean-pooled latent embedding

Edit the following variables in dna_sequence_analysis.py:

csv_path            = "<PATH_TO_YOUR_DATASET_CSV>"
sae_checkpoint_path = "<PATH_TO_YOUR_SAE_CHECKPOINT>"
output_path         = "<PATH_WHERE_YOU_WANT_MEANPOOL_WEIGHTS_CSV>"

Then run:

python dna_sequence_analysis.py