config.yaml

data_mix:
  (): dataset_construction.DataMix
  name: "code_30b"
  shuffle: false
  local_save_dir: data/
  load_from_local_save_dir: true
  compute_dataset_stats: true
  keep_separated_datasets_in_dataset_dict: true
  deduplicate_test_set: false
  ngram_path_for_extra_deduplication: null
  max_shard_size: "2GB"
  datasets:
# Code
    - (): dataset_construction.DatasetConfig
      dataset_path: 'bigcode/starcoderdata'
      train_split: "train[:20%]"
      text_column: "content"
      dataset_kwargs:
        data_dir: python
      build_test_set_from_train: true
    - (): dataset_construction.DatasetConfig
      dataset_path: 'bigcode/starcoderdata'
      train_split: "train[:20%]"
      text_column: "content"
      dataset_kwargs:
        data_dir: markdown
      build_test_set_from_train: true
    - (): dataset_construction.DatasetConfig
      dataset_path: 'bigcode/starcoderdata'
      train_split: "train[:20%]"
      text_column: "content"
      dataset_kwargs:
        data_dir: jupyter-scripts-dedup-filtered
      build_test_set_from_train: true
    - (): dataset_construction.DatasetConfig
      dataset_path: 'bigcode/starcoderdata'
      train_split: "train[:20%]"
      text_column: "content"
      dataset_kwargs:
        data_dir: jupyter-structured-clean-dedup
      build_test_set_from_train: true
    - (): dataset_construction.DatasetConfig
      dataset_path: 'bigcode/starcoderdata'
      train_split: "train[:15%]"
      text_column: "content"
      dataset_kwargs:
        data_dir: json
      build_test_set_from_train: true
    - (): dataset_construction.DatasetConfig
      dataset_path: 'teven/code_contests'
      train_split: "train"
      # build_test_set_from_train: true
      build_test_set_from_train: false
      test_split: "valid"
      # needs to be debugged
      filtering_function:
        (): dataset_collection.code.deepmind_code_contest.CodeContestFilter
      preprocessing_function:
        (): dataset_collection.code.deepmind_code_contest.CodeContestMapper
    - (): dataset_construction.DatasetConfig
      dataset_path: 'vikp/pypi_clean'
      train_split: "train"
      build_test_set_from_train: true
      text_column: code
      id_column: path
    # we could add synthetic textbook datasets or code with explanations !


# tokenizer: "mistralai/Mistral-7B-v0.1"
tokenizer: "manu/tok-fr-en-code"