hydra_config_1713758125.4750285.yaml

experiment:
  seed: 42
  name: suchir-demo
  group: climb
  dry_run: false
  offline_run: false
  resume_checkpoint_path: null
  resume_run_id: null
dataset:
  name: cambridge-climb/BabyLM
  subconfig: strict_small
tokenizer:
  name: cambridge-climb/CamBabyTokenizer-8192
  add_prefix_space: true
data_preprocessing:
  include_punctuation: true
  join_sentences: true
  max_input_length: 128
  callback_functions: null
model:
  name: roberta_pre_layer_norm
  model_kwargs:
    vocab_size: 8192
    num_hidden_layers: 8
    num_attention_heads: 8
    hidden_size: 256
    intermediate_size: 2048
    layer_norm_eps: 1.0e-05
    eos_token_id: 4
    bos_token_id: 3
    pad_token_id: 1
    tie_word_embeddings: false
trainer:
  batch_size: 32
  lr: 0.001
  num_warmup_steps: 100000
  max_training_steps: 400000
  eval_blimp: true
  eval_glue: false
  eval_msgs: false
  eval_perplexity: true
objective_curriculum:
  units:
    mlm:
      task_head_params: {}
      optimizer_params:
        lr: 0.001
      scheduler_params: {}
      optional_kwargs:
        mask_probability: 0.15
        unmask_probability: 0
  steps:
    mlm:
    - 0.0
    - 1.0
data_curriculum: null
vocabulary_curriculum: null