LeTI: Learning to Generate from Textual Interactions

Official model checkpoint for paper LeTI: Learning to Generate from Textual Interactions.

The code associated with these checkpoints can be found here.

Decompress the model checkpoint data.tar.gz (using tar -xzvf data.tar.gz) will create a folder data. After decompress the checkpoints, you can set export GS_BUCKET_PREFIX=pwd and run evaluation of your choice following this.

We currently release the following checkpoints in the paper:

• LeTI (350M): mbpp-ft/actor-rw-conditioned/codegen-350M-mono/350M+rw_conditioned+mixpretrain+50x3+lr1e-5/checkpoint_31416
• LeTI (2B): mbpp-ft/actor-rw-conditioned/codegen-2B-mono/2B+rw_conditioned+mixpretrain+10x3+lr5e-6/checkpoint_13465