checkpoints:
  checkpoint_interval: 1000
  checkpoints_path: /scratch/loubna/checkpoints/smollm-big-run-1p81G-smollm-1.7B-8T-seed-0-
  checkpoints_path_is_shared_file_system: false
  overwrite_datastage: false
  resume_checkpoint_path: /scratch/loubna/checkpoints/smollm-big-run-1p81G-smollm-1.7B-8T-seed-0-/3074000
  save_initial_state: true
data_stages:
- data:
    dataset:
      dataloader_type: cyclic
      dataset_max_tokens: null
      dataset_weights:
      - 1
      datasets:
      - filename_pattern: .*\.ds$
        folder: /fsx/loubna/tokenized_for_exps/phase2_mixture
        seed: 0
        shuffle: true
        skip_tokens: 0
      pad_samples_to_global_batch_size: false
      skip_in_stream: true
    num_loading_workers: 0
    seed: 0
  name: stable
  start_training_step: 1
experiment_logger:
  tensorboard_logger:
    push_to_hub_interval: 50
    repo_id: HuggingFaceTB/smollm-big-run
    repo_public: false
    tensorboard_dir: /scratch/loubna/tensorboard-cosmo-smollm-big-run
  wandb_logger:
    wandb_entity: loubnabnl
    wandb_project: smollm-big-run
general:
  benchmark_csv_path: null
  consumed_train_samples: 3927040000
  ignore_sanity_checks: true
  project: smollm-big-run
  run: smollm-big-run-1p81G-smollm-1.7B-8T-seed-0-
  seed: 42
  step: 3835000
kill_switch_path: null
lighteval:
  batch_size: 16
  checkpoints_path: null
  generation: null
  logging:
    hub_repo_details: null
    hub_repo_results: HuggingFaceTB/smollm-big-run
    hub_repo_tensorboard: HuggingFaceTB/smollm-big-run
    local_output_path: /scratch/loubna/lighteval/smollm-big-run-1p81G-smollm-1.7B-8T-seed-0-
    push_details_to_hub: false
    push_results_to_hub: true
    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/loubna/logs/smollmv2/eval-scripts
  slurm_template: /fsx/loubna/projects/brrr/examples/loubna/eval_1b.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:
    wandb_entity: loubnabnl
    wandb_project: smollm-big-run
    wandb_run_name: smollm-big-run-1p81G-smollm-1.7B-8T-seed-0-_evals
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: 0
    eos_token_id: 0
    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: 49152
optimizer:
  accumulate_grad_in_fp32: true
  clip_grad: 1.0
  learning_rate_scheduler:
    learning_rate: 0.0005
    lr_decay_starting_step: 3870000
    lr_decay_steps: 430000
    lr_decay_style: linear
    lr_warmup_steps: 2000
    lr_warmup_style: linear
    min_decay_lr: 0
  optimizer_factory:
    adam_beta1: 0.9
    adam_beta2: 0.95
    adam_eps: 1.0e-08
    name: adamW
    torch_adam_is_fused: true
  weight_decay: 0.01
  zero_stage: 0
parallelism:
  dp: 256
  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: /admin/home/loubna/miniconda3/envs/nanotron/bin/s5cmd
  upload_s3_path: s3://synthetic-project-models/big-run-5T/smollm-big-run-1p81G-smollm-1.7B-8T-seed-0-
tokenizer:
  tokenizer_max_length: null
  tokenizer_name_or_path: HuggingFaceTB/cosmo2-tokenizer
  tokenizer_revision: null
tokens:
  batch_accumulation_per_replica: 1
  limit_test_batches: 0
  limit_val_batches: 0
  micro_batch_size: 4
  sequence_length: 2048
  train_steps: 4300000
  val_check_interval: 100