/learning-dynamics-clm

Project repository for my Master's Thesis Investigating the Learning Dynamics of Conditional Language Models

Primary LanguagePython

Investigating the Learning Dynamics of Conditional Language Models

This is the project repository of my Master's Thesis on: Investigating the Learning Dynamics of Conditional Language Models

Model training

The scripts for model trianing can be found in the /scripts folder. Training scripts for specific model configurations are found in /models/tl.

Scripts for data preprocessing can be found in /data/tl.

Model translations

Translations have been generated using /compt/bleu/generate_test.sh.

KL divergence

To compute the KL divergence first install Fairseq provided in /alti/fairseq. KL divergence can be computed using /kl/test_on_time.sh.

ALTI+

Scirpts for alti+ computation are provided in alti/transformer-contribuions-nmt-v2. Run main.py to comptue the alti contirbuions over the course of training.

LRP

Our implementation of LRP in Fairseq can be found in /lrp/lrp_fairseq.

Hallucination metrics

LaBSE

The python script we used to compute LaBSE cosine similairty can be found at /comp/hallucinations/labse/labse.py.

Token hallucination metric

For the model-based token hallucination metric, clone and install the repository from the project repository. Download the pretrained XSum model provided. We used /halu/fairseq-detect-hallucination/test/run_test.sh for the computation of the token hallucination ratio.