Skip to content

Latest commit

 

History

History
83 lines (51 loc) · 3.11 KB

TRAIN.md

File metadata and controls

83 lines (51 loc) · 3.11 KB
Raw

Training

Setup

You should setup the following environment variable before executing the training script.

export GCLOUD_PROJECT=your-gcloud-project
export GS_BUCKET_PREFIX=gs://your-gs-bucket/your-directory

LeTI Training

Generally, each iteration (3 epochs) of LeTI takes about 22 hours for a 2B model, and 4 hours for 350M model on a TPU v3-8 VM instance.

Experiment on MBPP

Execute the corresponding script to train the model:

  • LeTI (350M): leti/scripts/model/train/rw_conditioned/mbpp/350M/finetune+50x3+lr1e-5+mixpretrain.sh

  • LeTI (350M) w/o textual feedback: leti/scripts/model/train/rw_conditioned/mbpp/350M/finetune+coarse-only+50x3+lr1e-5+mixpretrain.sh

  • LeTI (2B): leti/scripts/model/train/rw_conditioned/mbpp/2B/finetune+10x3+lr5e-6+mixpretrain.sh

  • LeTI (2B) w/o textual feedback: leti/scripts/model/train/rw_conditioned/mbpp/2B/finetune+coarse-only+30x3+lr5e-6+mixpretrain.sh

Ablations on MBPP

Changing the number of training problems |P|

You can compare the following results to baseline that uses full dataset (i.e., LeTI 350M or 2B).

2B

  • |P| = 16: leti/scripts/model/train/rw_conditioned/mbpp/2B/finetune+10x3+lr5e-6+mixpretrain+d16.sh
  • |P| = 64: leti/scripts/model/train/rw_conditioned/mbpp/2B/finetune+10x3+lr5e-6+mixpretrain+d64.sh
  • |P| = 128: leti/scripts/model/train/rw_conditioned/mbpp/2B/finetune+10x3+lr5e-6+mixpretrain+d128.sh

350M

  • |P| = 16: leti/scripts/model/train/rw_conditioned/mbpp/350M/finetune+50x3+lr1e-5+mixpretrain+d16.sh
  • |P| = 64: leti/scripts/model/train/rw_conditioned/mbpp/350M/finetune+50x3+lr1e-5+mixpretrain+d64.sh
  • |P| = 128: leti/scripts/model/train/rw_conditioned/mbpp/350M/finetune+50x3+lr1e-5+mixpretrain+d128.sh

Changing the number of sampled solution n per problem

You can compare the following results to baseline that uses n=128 (i.e., LeTI 350M or 2B).

2B

  • n=16: leti/scripts/model/train/rw_conditioned/mbpp/2B/finetune+10x3+lr5e-6+mixpretrain+n16.sh
  • n=64: leti/scripts/model/train/rw_conditioned/mbpp/2B/finetune+10x3+lr5e-6+mixpretrain+n64.sh

350M

  • n=16: leti/scripts/model/train/rw_conditioned/mbpp/350M/finetune+50x3+lr1e-5+mixpretrain+n16.sh
  • n=64: leti/scripts/model/train/rw_conditioned/mbpp/350M/finetune+50x3+lr1e-5+mixpretrain+n64.sh

Remove Continued Pre-training Regularization

2B leti/scripts/model/train/rw_conditioned/mbpp/2B/finetune+10x3+lr5e-6.sh

350M leti/scripts/model/train/rw_conditioned/mbpp/350M/finetune+50x3+lr1e-5.sh

Only do Continued Pre-training

2B leti/scripts/model/train/rw_conditioned/mbpp/2B/finetune+10x3+lr5e-6+onlypretrain.sh

350M leti/scripts/model/train/rw_conditioned/mbpp/350M/finetune+50x3+lr1e-5+onlypretrain.sh

Experiment on Event Argument Extraction (EAE)

You need to start an API server before start the training script using leti/verifier/eae/run_api_server.sh.

leti/scripts/model/train/rw_conditioned/eae/350M/finetune+30x3+lr1e-5+n64.sh

Finetuned baselines for comparison

MBPP Finetuning

2B leti/scripts/model/train/actor/mbpp/jax-2B.sh

350M leti/scripts/model/train/actor/mbpp/jax-350M.sh