Update big.yml

big.yml

@@ -27,7 +27,7 @@ datasets:
 
 shuffle_merged_datasets: true
 dataset_prepared_path: prepared_data
-val_set_size: 0.0
+val_set_size: 0.01
 output_dir: ./output
 
 plugins:
@@ -43,7 +43,7 @@ cut_cross_entropy: true
 #torch_compile: auto
 #torch_compile_backend:
 
-sequence_len: 32768
+sequence_len: 16384
 sample_packing: true
 eval_sample_packing: false
 pad_to_sequence_len: true
@@ -80,7 +80,7 @@ bf16: auto
 fp16:
 tf32: false
 
-gradient_checkpointing: unsloth
+gradient_checkpointing: true
 early_stopping_patience:
 resume_from_checkpoint:
 local_rank:
@@ -90,9 +90,9 @@ flash_attention: true
 s2_attention:
 
 warmup_steps: 40
-evals_per_epoch:
+evals_per_epoch: 2
 eval_table_size:
-eval_max_new_tokens:
+eval_max_new_tokens: 128
 saves_per_epoch: 2
 debug:
 deepspeed: ./deepspeed_configs/zero3_bf16_compile.json