text/pretraining/smollm1/config_smollm1_1B.yaml

# SmolLM1 1.7B trained on 1T tokens
checkpoints:
  checkpoint_interval: 2000
  checkpoints_path: checkpoints
  checkpoints_path_is_shared_file_system: false
  resume_checkpoint_path: null
  save_final_state: false
  save_initial_state: false
data_stages:
- data:
    dataset:
      dataset_folder: # paths to tokenized datasets
        - datasets/fineweb-edu-dedup
        - datasets/cosmopedia-v2
        - datasets/open-web-math
        - datasets/starcoderdata-python
        - datasets/stackoverflow
      dataset_weights:
        - 0.7
        - 0.15
        - 0.06
        - 0.08
        - 0.01
    num_loading_workers: 1
    seed: 42
  name: training stage
  start_training_step: 1
- data:
    dataset: # we change data mixture to use python-edu
      dataset_folder: 
        - datasets/fineweb-edu-dedup
        - datasets/cosmopedia-v2
        - datasets/open-web-math
        - datasets/python-edu
        - datasets/stackoverflow
        - datasets/deepmind_mathematics
      dataset_weights:
        - 0.7
        - 0.15
        - 0.055
        - 0.08
        - 0.01
        - 0.005
    num_loading_workers: 1
    seed: 42
  name: training stage 2
  start_training_step: 300000
general:
  benchmark_csv_path: null
  consumed_train_samples: null
  ignore_sanity_checks: true
  project: smollm
  run: smollm-1700M
  seed: 8
  step: 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: 0.022097086912079608
  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
    rope_theta: 10000.0
    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: 400000
    lr_decay_steps: 100000
    lr_decay_style: 1-sqrt
    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: 64 # 8 nodes
  expert_parallel_size: 1
  pp: 1
  pp_engine: 1f1b
  recompute_layer: false
  tp: 1
  tp_linear_async_communication: true
  tp_mode: REDUCE_SCATTER
  tp_recompute_allgather: true
profiler: null
tokenizer:
  tokenizer_max_length: null
  tokenizer_name_or_path: HuggingFaceTB/cosmo2-tokenizer
  tokenizer_revision: null
tokens:
  batch_accumulation_per_replica: 4
  limit_test_batches: 0
  limit_val_batches: 0
  micro_batch_size: 4 # GBS = 4*4*64*sequence_length = 1024*sequence_length = 2.1M tokens
  sequence_length: 2048
  train_steps: 500000
  val_check_interval: -1