Upload run root files (non-recursive) - training_config.yaml

training_config.yaml

@@ -0,0 +1,32 @@
+bf16: true
+cutoff_len: 4096
+dataset: tulu3_mixture_coding
+ddp_timeout: 180000000
+deepspeed: examples/deepspeed/ds_z3_config.json
+do_train: true
+eval_steps: 100
+eval_strategy: steps
+finetuning_type: full
+gradient_accumulation_steps: 32
+learning_rate: 5.0e-06
+logging_steps: 10
+lr_scheduler_kwargs:
+  cooldown_ratio: 0.4
+lr_scheduler_type: warmup_stable_decay
+model_name_or_path: meta-llama/Llama-3.1-8B
+num_train_epochs: 1.0
+output_dir: /scratch/pxm5426/runs/lora-exploration/llama-factory/Llama-3.1-8B_tulu3_mixture_coding_full_adamw_ebs128_lr5e-06_wsd-cr0.4
+overwrite_cache: true
+overwrite_output_dir: true
+per_device_eval_batch_size: 2
+per_device_train_batch_size: 2
+plot_loss: true
+preprocessing_num_workers: 16
+report_to: wandb
+save_steps: 1200
+save_strategy: steps
+stage: sft
+template: tulu_v3
+trust_remote_code: true
+val_size: 0.01
+warmup_ratio: 0.0