config_stage_3.yaml

# Directory settings
checkpoint_dir: "/lustre/scratch/data/polyglot_datasets/portuguese/checkpoints/models/tucano_v2"
train_dataset_dir:
  # Total: ~78B
  # Portuguese Text (~58B, 75%) | English Text (~20B, 25%)
  # Web Text (~28B, 35%)
  - /lustre/scratch/data/polyglot_datasets/portuguese/tokenized/gigaverbo_v2/4 # 28B (PT)
  # Synthetic Text (~38B, %)
  - /lustre/scratch/data/polyglot_datasets/portuguese/tokenized/gigaverbo_v2_synth # 10B (PT)
  - /lustre/scratch/data/polyglot_datasets/portuguese/tokenized/gigaverbo_v2_synth # 10B (PT)
  - /lustre/scratch/data/polyglot_datasets/portuguese/tokenized/gigaverbo_v2_synth # 10B (PT)
  # Reasoning Text (~12B, 15%)
  - /lustre/scratch/data/polyglot_datasets/portuguese/tokenized/nvidia_openscience # 9B (EN)
  - /lustre/scratch/data/polyglot_datasets/portuguese/tokenized/big_reasoning_traces # 2B (EN)
  - /lustre/scratch/data/polyglot_datasets/portuguese/tokenized/math_meta_reasoning_filtered # 1B (EN)
  # Math Text (~8B, 10%)
  - /lustre/scratch/data/polyglot_datasets/portuguese/tokenized/finemath/4 # 8B (EN)
val_dataset_dir: "/lustre/scratch/data/polyglot_datasets/portuguese/tokenized/validation"
dataset_type: "parquet"
cache_dir: "/lustre/mlnvme/data/polyglot/.cache"

# Data loading settings
pin_memory: true
num_workers_for_dataloader: 16
shuffle_dataset: true
mask_eos_token: false
mask_pad_token: false

# Model architecture settings
vocab_size: 49152
num_hidden_layers: 28
num_attention_heads: 16
num_key_value_heads: 8
head_dim: null
hidden_size: 1536
intermediate_size: 3072
max_position_embeddings: 4096
tie_word_embeddings: true
hidden_act: "silu"
output_hidden_states: false
attn_implementation: "flash_attention_2"
use_cache: false
no_rope_layer_interval: null
rope_theta: 50000.0
rope_scale_factor: null
rms_norm_eps: 0.000001

# Training settings
total_batch_size: 2097152
micro_batch_size: 16
eval_micro_batch_size: 8
num_train_epochs: 1
warmup_steps: 0
max_learning_rate: 0.0007
min_learning_rate: 0.0
muon_learning_rate: 0.007
weight_decay: 0.1
beta1: 0.9
beta2: 0.95
eps: 0.00000001
lr_decay_type: "wsd"
use_sqrt: true
lr_decay_iters_coef: 1.
seed: 42
max_steps: 35000
max_grad_norm: 1.0

# Precision and optimization settings
torch_compile: false
mat_mul_precision: "highest"
tf32: true
bf16: true
gradient_checkpointing: false
use_liger_kernel: true
static_graph: false

# Hub settings
push_to_hub: false
hub_token: null
hub_model_id: null

# Tokenizer and Reference model
tokenizer_name_or_path: "/lustre/scratch/data/polyglot_datasets/portuguese/checkpoints/tokenizers/sentencepiece"
chat_template_path: null
reference_model: "HuggingFaceTB/SmolLM2-360M"
continual_pretraining: false

# Checkpoint settings
resume_from_checkpoint: null
checkpointing_steps: 5000
begin_new_stage: false
stage_name: "stage3"

# Miscellaneous settings
sanity_check: false
sanity_check_num_samples: 100000
wandb_token: null
wandb_id: "tucano-v2"
wandb_project: "Polyglot"
wandb_desc: "Developing LLMs for low-resource languages"