# Directory settings
checkpoint_dir: "/lustre/scratch/data/polyglot_datasets/portuguese/checkpoints/models/Tucano2-qwen-0.5B-Base"
train_dataset_dir:
  # Total: ~48B
  # Web Text (~28B, 58%)
  - /lustre/scratch/data/polyglot_datasets/portuguese/tokenized/gigaverbo_v2/5 # 0.1B (PT)
  - /lustre/scratch/data/polyglot_datasets/portuguese/tokenized/gigaverbo_v2/4 # 28B (PT)
  # Synthetic Text (~20B, 42%)
  - /lustre/scratch/data/polyglot_datasets/portuguese/tokenized/gigaverbo_v2_synth # 10B (PT)
  - /lustre/scratch/data/polyglot_datasets/portuguese/tokenized/gigaverbo_v2_synth # 10B (PT)
val_dataset_dir: "/lustre/scratch/data/polyglot_datasets/portuguese/tokenized/validation"
dataset_type: "parquet"
cache_dir: "/lustre/mlnvme/data/nanotronics/.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: 128
hidden_size: 1024
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: 1000000.0
rope_scale_factor: null
rms_norm_eps: 0.000001

# Training settings
total_batch_size: 1048576
micro_batch_size: 8
eval_micro_batch_size: 8
num_train_epochs: 1
warmup_steps: 100
max_learning_rate: 0.00025
min_learning_rate: 0.0
muon_learning_rate: null
weight_decay: 0.1
beta1: 0.9
beta2: 0.95
eps: 0.00000001
lr_decay_type: "cosine"
use_sqrt: true
lr_decay_iters_coef: 1.
seed: 42
max_steps: 50000
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/models/tucano2-qwen-0.5B"
chat_template_path: null
reference_model: "/lustre/scratch/data/polyglot_datasets/portuguese/checkpoints/models/tucano2-qwen-0.5B"
continual_pretraining: true

# Checkpoint settings
resume_from_checkpoint: null
checkpointing_steps: 2500
begin_new_stage: true
stage_name: "single_cosine"

# Miscellaneous settings
sanity_check: false
sanity_check_num_samples: 100000
wandb_token: null
wandb_id: "tucano2-qwen-0.5b"
wandb_project: "Polyglot"
wandb_desc: "Developing LLMs for low-resource languages"