mix-fw_edu_hq-article_granite3_1b_hq/4000/config.yaml

checkpoints:
  checkpoint_interval: 500
  checkpoints_path: /scratch/joel_niklaus/tmp/finephrase/checkpoints/mix-fw_edu_hq-article_granite3_1b_hq
  checkpoints_path_is_shared_file_system: false
  load_lr_scheduler: true
  load_optimizer: true
  resume_checkpoint_path: null
  save_final_state: true
  save_initial_state: false
data_stages:
- data:
    dataset:
      dataset_folder:
      - /scratch/joel_niklaus/tmp/finephrase/dataset/fw_edu_hq/
      - /scratch/joel_niklaus/tmp/finephrase/dataset/article_granite3_1b_hq/
      dataset_max_tokens: null
      dataset_read_path: null
      dataset_weights:
      - 0.5
      - 0.5
      pad_samples_to_global_batch_size: false
      return_positions: true
      shuffle_files: false
      skip_in_stream: false
      token_size_in_bytes: 4
      tokenizer_name: hynky/Llama-3.2-1B-no-bos
      use_old_brrr_dataloader: false
      vocab_size: 128256
    num_loading_workers: 0
    seed: 6
  name: stable
  start_training_step: 1
general:
  benchmark_csv_path: null
  consumed_train_samples: 2048000
  ignore_sanity_checks: true
  project: finephrase
  run: mix-fw_edu_hq-article_granite3_1b_hq
  seed: 6
  step: 4000
lighteval:
  batch_size: 8
  eval_interval: 500
  eval_interval_file: null
  local_checkpoint_dir: /scratch/joel_niklaus/tmp/finephrase/evals-ckpt
  logs_path: /fsx/joel_niklaus/logs/finephrase/experiments/evals
  nanotron_path: /fsx/joel_niklaus/projects/finephrase/nanotron
  output_dir: s3://finephrase/experiments/evals-test
  parallelism:
    context_parallel_size: 1
    dp: 1
    expert_parallel_size: 1
    moe_layer_recompute: false
    pp: 1
    pp_engine: afab
    recompute_layer: false
    tp: 1
    tp_linear_async_communication: true
    tp_mode: ALL_REDUCE
    tp_recompute_allgather: true
  s3_save_path: null
  slurm:
    cpus_per_task: 88
    gpus_per_node: 8
    hf_cache: /fsx/joel_niklaus/.cache/huggingface
    partition: hopper-prod
    qos: normal
    reservation: null
    time: '1:00:00'
  tasks:
    custom_tasks: /fsx/joel_niklaus/projects/finephrase/task_list.py
    dataset_loading_processes: 8
    max_samples: 1000
    multichoice_continuations_start_space: null
    no_multichoice_continuations_start_space: null
    num_fewshot_seeds: null
    tasks: /fsx/joel_niklaus/projects/finephrase/tasks.txt
  upload_to_wandb: false
  wandb: null
  wandb_entity: null
  wandb_project: null
logging:
  iteration_step_info_interval: 5
  log_level: info
  log_level_replica: info
metrics_logging:
  log_detail_interval: 200
  log_level: 1
model:
  ddp_bucket_cap_mb: 50
  dtype: bfloat16
  init_method:
    scaling_method: NUM_LAYERS
    std: 0.02
  make_vocab_size_divisible_by: 1
  model_config:
    _attn_implementation: flash_attention_2
    _fused_rms_norm: true
    _fused_rotary_emb: true
    _use_doc_masking: true
    _use_qkv_packed: true
    attention_bias: false
    bos_token_id: 128000
    eos_token_id: 128001
    flex_attention_mask: null
    hidden_act: silu
    hidden_size: 2048
    initializer_range: 0.02
    intermediate_size: 6144
    is_qwen2_config: true
    max_position_embeddings: 4096
    moe_config: null
    no_rope_layer: null
    num_attention_heads: 16
    num_hidden_layers: 28
    num_key_value_heads: 8
    pad_token_id: null
    pretraining_tp: 1
    rms_norm_eps: 1.0e-06
    rope_interleaved: false
    rope_scaling: null
    rope_theta: 10000
    sliding_window_size: null
    tie_word_embeddings: true
    use_cache: true
    vocab_size: 128256
    z_loss_coefficient: 1.0e-05
    z_loss_enabled: false
optimizer:
  accumulate_grad_in_fp32: true
  clip_grad: 1.0
  learning_rate_scheduler:
    learning_rate: 0.0005
    lr_decay_starting_step: 9000
    lr_decay_steps: 1000
    lr_decay_style: linear
    lr_warmup_steps: 100
    lr_warmup_style: linear
    min_decay_lr: 5.0e-05
  optimizer_factory:
    adam_beta1: 0.9
    adam_beta2: 0.95
    adam_eps: 1.0e-08
    name: adamW
    torch_adam_is_fused: true
  weight_decay: 0.1
  weight_decay_exclude_named_params:
  - .*token_embedding.*
  zero_stage: 0
parallelism:
  context_parallel_size: 1
  dp: 64
  expert_parallel_size: 1
  moe_layer_recompute: false
  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
s3_upload:
  remove_after_upload: true
  s5cmd_concurrency: 10
  s5cmd_numworkers: 32
  s5cmd_path: /fsx/joel_niklaus/projects/finephrase/.venv/bin/s5cmd
  upload_s3_path: s3://finephrase/experiments/checkpoints/mix-fw_edu_hq-article_granite3_1b_hq
tokenizer:
  tokenizer_max_length: 4096
  tokenizer_name_or_path: hynky/Llama-3.2-1B-no-bos
  tokenizer_revision: null
tokens:
  batch_accumulation_per_replica: 4
  limit_test_batches: 0
  limit_val_batches: 0
  micro_batch_size: 2
  sequence_length: 4096
  train_steps: 10000
  val_check_interval: 0