model_likelihood.yaml

defaults:
  - _self_
  - hydra: default
  - data: pubmedqa

model:
  name_or_path: "meta-llama/Meta-Llama-3-8B"
  source: hf
  temperature: 1.0
  top_k: -1
  device: 0
  target_distribution: "y|x"
  precision: "bf16"
  tokenizer:
    name_or_path:
    source:
    add_pad_token: True

generator:
  name_or_path: cemde/Domain-Certification-MedQA-Guide-Finetuned
  source: hf
  temperature: 1.0
  top_k: -1
  device: 0
  target_distribution: "y"
  precision: "none"
  tokenizer:
    name_or_path:
    source:
    add_pad_token: False

tokenizers_match: False # True if the tokenizers for models $L$ and $G$ are functionally the same, False otherwise.

run:
  seed: 23633
  compile: True

inference:
  task: "causal" # "seq2seq" or "causal"
  batch_size: 16
  prompt_length: "dataset" # can be set to "<int>" or "dataset". In the first case, the query+response will be cut after <int> tokens. Otherwise the natural split of Q and A will be used.
  drop_last_batch: True # should be True. False can lead to some errors when concatenating the likelihoods.
  shuffle_batches: False # should be False for reproducibility. Can however bias the dataset if not previously shuffled and drop_last_batch is True.

data_config_name: None # DO NOT SET. Will be dome automatically

log:
  print_to_file: False


# documentation:
# If ...tokenizer.name_or_path is None, name_or_path and source will be set to the same value as the model / generator.