config_stage_1.yaml

# Directory settings
checkpoint_dir: "/lustre/scratch/data/polyglot_datasets/bengali/checkpoints/models/LilTii/v2"
train_dataset_dir:
  # Total: ~99B
  # Bengali Text (~40B)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/edu_score_1"                 # 5.8B (ben)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/edu_score_2"                 # 8.6B (ben)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/edu_score_3"                 # 4.2B (ben)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/edu_score_4"                 # 1.5B (ben)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/edu_score_5"                 # 5.5M (ben)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/edu_score_1"                 # 5.8B (ben)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/edu_score_2"                 # 8.6B (ben)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/edu_score_3"                 # 4.2B (ben)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/edu_score_4"                 # 1.5B (ben)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/edu_score_5"                 # 5.5M (ben)
  # Edu English Text (~35B)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/fineweb_edu/edu_score_3"        # 35.B (eng)
  # Reasoning (~14.6B)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/math_meta_reasoning_filtered"   # 1.2B (eng)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/math_meta_reasoning_filtered"   # 1.2B (eng)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/nvidia_openscience"             # 9.8B (eng)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/big_reasoning_traces"           # 2.4B (eng)
  # Edu Math Text (~9.5B)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/finemath_34b/edu_score_4"       # 8.5B (eng)
  - "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/finemath_34b/edu_score_5"       # 1.0B (eng)
val_dataset_dir: "/lustre/scratch/data/polyglot_datasets/bengali/tokenized/validation_split"  
dataset_type: "parquet"
cache_dir: "/lustre/mlnvme/data/polyglot/.cache"

# Data loading settings
pin_memory: true
num_workers_for_dataloader: 32
shuffle_dataset: true

# 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: 2000
max_learning_rate: 0.0007
min_learning_rate: 0.0
weight_decay: 0.1
beta1: 0.9
beta2: 0.95
eps: 0.00000001
lr_decay_type: "wsd"
lr_decay_iters_coef: 0.0
seed: 1337
max_steps: 47500
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: "Polygl0t/LilTii-v0.2"
reference_model: "HuggingFaceTB/SmolLM2-360M"

# Checkpoint settings
resume_from_checkpoint: null
checkpointing_steps: 2500
begin_new_stage: false
stage_name: "Warmup-Stable"

# Miscellaneous settings
sanity_check: false
sanity_check_num_samples: 100000
wandb_token: null
wandb_id: "LilTii-v0.2"
wandb_project: "Polyglot"
wandb_desc: "Developing LLMs for low-resource languages"