Upload configs.yaml with huggingface_hub

configs.yaml

@@ -1,34 +1,27 @@
-adam_beta1: 0.9
-adam_beta2: 0.999
 bf16: true
-cutoff_len: 2048
+cutoff_len: 16384
 dataset: mlfoundations-dev/r1_annotated_math
 dataset_dir: ONLINE
 ddp_timeout: 180000000
 deepspeed: dcft/train/zero3.json
 do_train: true
-enable_liger_kernel: false
-eval_strategy: epoch
+eval_strategy: 'no'
 finetuning_type: full
 formatting: sharegpt
-global_batch_size: 512
-gradient_accumulation_steps: 1
-gradient_checkpointing: true
+global_batch_size: 96
+gradient_accumulation_steps: 3
 hub_model_id: mlfoundations-dev/llama3-1_8b_r1_annotated_math
-include_hp: dcft/train/hp_settings/hritik.yaml
-learning_rate: 5.0e-06
-logging_steps: 10
-lr_scheduler_type: constant
-max_grad_norm: 1
+include_hp: dcft/train/hp_settings/reasoning.yaml
+learning_rate: 1.0e-05
+logging_steps: 1
+lr_scheduler_type: cosine
+max_samples: 1000000
 messages: conversations
-model_name_or_path: meta-llama/Meta-Llama-3.1-8B
-neat_packing: true
+model_name_or_path: Qwen/Qwen2.5-7B-Instruct
 num_train_epochs: 3.0
 output_dir: /tmp/dcft_checkpoints/llama3-1_8b_r1_annotated_math
 overwrite_cache: true
-overwrite_output_dir: true
-packing: true
-per_device_train_batch_size: 16
+per_device_train_batch_size: 1
 plot_loss: true
 preprocessing_num_workers: 16
 push_to_db: true
@@ -37,6 +30,5 @@ report_to: wandb
 run_name: llama3-1_8b_r1_annotated_math
 save_strategy: epoch
 stage: sft
-template: llama3
-val_size: 0.05
-weight_decay: 0.1
+template: qwen25
+warmup_ratio: 0.1