assistant_tag: assistant
bf16: true
content_tag: value
cutoff_len: 16384
dataset: /leonardo_work/EUHPC_E03_068/marianna/lf_datasets/mlfoundations-dev/extra_verified
dataset_dir: ONLINE
ddp_timeout: 180000000
deepspeed: dcft/train/zero3_offload.json
do_train: true
enable_liger_kernel: true
finetuning_type: full
formatting: sharegpt
global_batch_size: 384
gradient_accumulation_steps: 1
hub_model_id: mlfoundations-dev/mlfoundations-dev_extra_verified-32B
learning_rate: 1.0e-05
logging_steps: 1
lr_scheduler_type: cosine
max_samples: 1000000
messages: conversations
model_name_or_path: Qwen/Qwen2.5-32B-Instruct
num_train_epochs: 3.0
output_dir: /leonardo_scratch/large/userexternal/mnezhuri/dcft_checkpoints/saves/DCFT-Stratos-mlfoundations-dev_extra_verified-32B
overwrite_cache: true
per_device_train_batch_size: 1
plot_loss: true
preprocessing_num_workers: 16
push_to_db: false
push_to_hub: false
report_to: wandb
resume_from_checkpoint: /leonardo_scratch/large/userexternal/mnezhuri/dcft_checkpoints/saves/DCFT-Stratos-mlfoundations-dev_extra_verified-32B/checkpoint-850
role_tag: from
run_name: DCFT-Stratos-mlfoundations-dev_extra_verified-32B
save_steps: 50
stage: sft
template: qwen25
tokenized_path: /leonardo_work/EUHPC_E03_068/marianna/lf_datasets/mlfoundations-dev_extra_verified_tokenized
user_tag: user
warmup_ratio: 0.1