project to build the cross-encoder for the
recommender_systems and retrieval projects.

not finished testing yet.

This project is to fine-tune a pretrained re-ranker
model using the movie-lens dataset and a listwise
loss function.  
The hard-negative lists for training, validation
and tests were made in the project:
   https://github.com/nking/retrieval.git

The pretrained model used here is:
    castorini/LiT5-Distill-base-v2
see https://huggingface.co/castorini/LiT5-Distill-base-v2
see https://github.com/castorini/LiT5
and their paper:
 [2312.16098] Scaling Down, LiTting Up: Efficient Zero-Shot Listwise Reranking with Seq2seq Encoder-Decoder Models
@ARTICLE{tamber2023scaling,
  title   = {Scaling Down, LiTting Up: Efficient Zero-Shot Listwise Reranking with Seq2seq Encoder-Decoder Models},
  author  = {Manveer Singh Tamber and Ronak Pradeep and Jimmy Lin},
  year    = {2023},
  journal = {arXiv preprint arXiv: 2312.16098}
}

The T5-based sequence-to-sequence encoder-decoder model was
designed for efficient zero-shot listwise reranking.
Tambor, Pradeep, and Lin  provide a distilled version with
a 200 million parameters.

gemini.google.com was especially useful in finding pre-trainined
models and in changes needed for distributed training.

=================================

to install dependencies:

  consider setting up a virtual environment

  pip install --editable . 

to run the code on a very small test dataset:

  cd to project base directory
 
  python src/test/python/movie_lens_reranker/fine_tune_cross_encoder_listwise_lit5_test.py