youtube/config.yaml

checkpoints:
  checkpoint_interval: 500
  checkpoints_path: /scratch/craffel/checkpoints/commav0p1-ablations-1p82G-commonpile0p1filteredyoutube-seed-6-
  checkpoints_path_is_shared_file_system: false
  resume_checkpoint_path: s3://comma-v0.1-ablations/checkpoints/commav0p1-ablations-1p82G-commonpile0p1filteredyoutube-seed-6-
  save_initial_state: true
data:
  dataset:
    dataloader_type: single
    dataset_max_tokens: null
    dataset_weights: null
    datasets:
    - bits_per_token: 16
      filename_pattern: .*\.ds$
      folder: /scratch/dataset/commav0p1-ablations-1p82G-commonpile0p1filteredyoutube-seed-6-/
      original_folder: null
      seed: 6
      shuffle: true
      skip_tokens: 0
    pad_samples_to_global_batch_size: false
    skip_in_stream: true
  num_loading_workers: 0
  seed: 6
experiment_logger:
  tensorboard_logger:
    push_to_hub_interval: 300
    repo_id: craffel/commav0p1-ablations
    repo_public: false
    tensorboard_dir: /scratch/craffel/tensorboard-craffel-commav0p1-ablations
  wandb_logger: null
general:
  benchmark_csv_path: null
  consumed_train_samples: 14336000
  ignore_sanity_checks: true
  project: commav0p1-ablations
  run: commav0p1-ablations-1p82G-commonpile0p1filteredyoutube-seed-6-
  seed: 42
  step: 14000
kill_switch_path: null
lighteval:
  batch_size: 16
  checkpoints_path: null
  generation: null
  logging:
    hub_repo_details: null
    hub_repo_results: null
    hub_repo_tensorboard: craffel/commav0p1-ablations
    local_output_path: /scratch/craffel/lighteval/commav0p1-ablations-1p82G-commonpile0p1filteredyoutube-seed-6-
    push_details_to_hub: false
    push_results_to_hub: false
    push_results_to_tensorboard: true
    tensorboard_metric_prefix: e
  parallelism:
    dp: 8
    expert_parallel_size: 1
    pp: 1
    pp_engine: 1f1b
    tp: 1
    tp_linear_async_communication: false
    tp_mode: ALL_REDUCE
  slurm_script_dir: /fsx/craffel/train/eval-scripts
  slurm_template: /fsx/craffel/run_eval.slurm.jinja
  tasks:
    custom_tasks: brrr.lighteval.evaluation_tasks
    dataset_loading_processes: 8
    max_samples: 1000
    multichoice_continuations_start_space: null
    no_multichoice_continuations_start_space: null
    num_fewshot_seeds: null
    tasks: early-signal
  wandb: null
logging:
  iteration_step_info_interval: 1
  log_level: info
  log_level_replica: info
model:
  ddp_bucket_cap_mb: 25
  dtype: bfloat16
  init_method:
    std: 0.02
  make_vocab_size_divisible_by: 1
  model_config:
    bos_token_id: 1
    eos_token_id: 2
    hidden_act: silu
    hidden_size: 2048
    initializer_range: 0.02
    intermediate_size: 8192
    is_llama_config: true
    max_position_embeddings: 2048
    num_attention_heads: 32
    num_hidden_layers: 24
    num_key_value_heads: 32
    pad_token_id: null
    pretraining_tp: 1
    rms_norm_eps: 1.0e-05
    rope_scaling: null
    tie_word_embeddings: true
    use_cache: true
    vocab_size: 50272
optimizer:
  accumulate_grad_in_fp32: true
  adam_beta1: 0.9
  adam_beta2: 0.95
  adam_eps: 1.0e-08
  clip_grad: 1.0
  learning_rate_scheduler:
    learning_rate: 0.0003
    lr_decay_starting_step: null
    lr_decay_steps: null
    lr_decay_style: cosine
    lr_warmup_steps: 500
    lr_warmup_style: linear
    min_decay_lr: 3.0e-05
  torch_adam_is_fused: true
  weight_decay: 0.1
  zero_stage: 0
parallelism:
  dp: 64
  expert_parallel_size: 1
  pp: 1
  pp_engine: 1f1b
  tp: 1
  tp_linear_async_communication: true
  tp_mode: REDUCE_SCATTER
profiler: null
s3_upload:
  remove_after_upload: true
  s5cmd_concurrency: 5
  s5cmd_numworkers: 16
  s5cmd_path: /fsx/craffel/miniconda3/envs/exp/bin/s5cmd
  upload_s3_path: s3://comma-v0.1-ablations/checkpoints/commav0p1-ablations-1p82G-commonpile0p1filteredyoutube-seed-6-
tokenizer:
  tokenizer_max_length: null
  tokenizer_name_or_path: gpt2
  tokenizer_revision: null
tokens:
  batch_accumulation_per_replica: 4
  limit_test_batches: 0
  limit_val_batches: 0
  micro_batch_size: 4
  sequence_length: 2048
  train_steps: 14305
  val_check_interval: 100