base_model: Qwen/Qwen3-14B-Base

trust_remote_code: true
tokenizer_use_fast: true

load_in_8bit: false
load_in_4bit: false


datasets:
  - path: philipperen55/datasetCPT70axolotlRandomized
    split: train
    data_files: datasetCPT70axolotlRandomized.jsonl
    type: completion
    field: text


val_set_size: 0.0001

dataset_prepared_path: prepared_cpt
output_dir: outputs_cpt

sequence_len: 2048
pad_to_sequence_len: true
sample_packing: true
eval_sample_packing: true
attn_implementation: flash_attention_2
excess_length_strategy: truncate
train_on_inputs: true 
add_eos_token: true

# FULL finetune CPT
micro_batch_size: 2
gradient_accumulation_steps: 8
num_epochs: 1


optimizer: adamw_8bit
learning_rate: 4e-5
weight_decay: 0.01
lr_scheduler: constant_with_warmup
warmup_ratio: 0.01
max_grad_norm: 1.0

fp16: false
bf16: true
tf32: true
gradient_checkpointing: false

logging_steps: 50
eval_steps: 1000
save_steps: 5000
save_total_limit: 1
save_only_model: true

seed: 42
#mettre 24 si ya plus de 24 vspu, sinon mettre 16 si ya 24vcpu
dataset_num_proc: 24

# WandB
wandb_project: qwen3_14b_cpt_full_axolotl

# Hub
hub_model_id:  
push_to_hub: false #mon script upload de façon plus sûre
hub_strategy: every_save