Skip to content

README_THESIS.md

Latest commit

 

History

History
76 lines (57 loc) · 2.1 KB

README_THESIS.md

File metadata and controls

76 lines (57 loc) · 2.1 KB

Training scripts and data

Data

  • Folder thesis_data includes all training files
  • Under thesis_data there are following main folders
ls /home/olli/UniHY/thesis_specter/thesis_data/training/ 
data.json  metadata.json  README.md  test.txt  train.txt  val.txt
ls -R /home/olli/UniHY/thesis_specter/thesis_data/inference/ 
inference:
models  results

inference/models:
finnish_bert_base_cased_v1  README.md  sbert-cased-finnish-paraphrase

inference/models/finnish_bert_base_cased_v1:
finnish_bert.tar.gz

inference/models/sbert-cased-finnish-paraphrase:

inference/results:

models files for baseline models results are results from baseline models prediction. These are later used in scidocs

ls /home/olli/UniHY/thesis_specter/thesis_data/preprocessed/ 
data-metrics.json  data-test.p  data-train.p  data-val.p

=> These are training files required for script 4-run-training-script-thesis.sh

Create training files

When creating training files you need to provide three files

  • metadata.json
  • data.json
  • train.txt, val.txt and test.txt

If you're creating only samples use filtered citation data but full metadata. It's important that the cited articles are found from the metadata file.

Demo version using only samples.
./3-create-training-files-thesis.sh --demo --install_packages

Full version install packages: 
./3-create-training-files-thesis.sh --install_packages

Full version 
./3-create-training-files-thesis.sh

Run training script

TODOs

  • In SPECTER the data/vocab/ dir contains tokens and other text files.
  • tokens.txt length is around 240k
  • Instead scibert vocab.txt length is around 30k => Why there is a mismatch.
  • Check that finnish_bert.tar.gz can be used as model weights
  • Check that correct vocabulary is used
  • tokens vs vocab.txt
  • 2023-04-17 07:22:11,419 - INFO - allennlp.common.params - dataset_reader.token_indexers.bert.do_lowercase = True => Check if do_lowercase should be False in case of finnish bert
Demo version
./4-run-training-script-thesis.sh --demo

Full version 
./4-run-training-script-thesis.sh