new grpo logic (#274)

huggingface · Feb 11, 2025 · 52aa875 · 52aa875
1 parent 82b2a65
commit 52aa875
Show file tree

Hide file tree

Showing 3 changed files with 8 additions and 8 deletions.
diff --git a/recipes/DeepSeek-R1-Distill-Qwen-7B/grpo/config_demo.yaml b/recipes/DeepSeek-R1-Distill-Qwen-7B/grpo/config_demo.yaml
@@ -31,12 +31,12 @@ lr_scheduler_type: cosine
 max_prompt_length: 512
 max_completion_length: 1024
 max_steps: -1
-num_generations: 2
+num_generations: 7
 num_train_epochs: 1
 output_dir: data/DeepSeek-R1-Distill-Qwen-7B-GRPO
 overwrite_output_dir: true
-per_device_eval_batch_size: 4
-per_device_train_batch_size: 2
+per_device_eval_batch_size: 32
+per_device_train_batch_size: 16
 push_to_hub: true
 report_to:
 - wandb

diff --git a/recipes/Qwen2.5-1.5B-Instruct/grpo/config_demo.yaml b/recipes/Qwen2.5-1.5B-Instruct/grpo/config_demo.yaml
@@ -33,12 +33,12 @@ lr_scheduler_type: cosine
 max_prompt_length: 512
 max_completion_length: 1024
 max_steps: -1
-num_generations: 2
+num_generations: 7
 num_train_epochs: 1
 output_dir: data/Qwen2.5-1.5B-Open-R1-GRPO
 overwrite_output_dir: true
-per_device_eval_batch_size: 4   
-per_device_train_batch_size: 2
+per_device_eval_batch_size: 32
+per_device_train_batch_size: 16
 push_to_hub: true
 report_to:
 - wandb

diff --git a/recipes/Qwen2.5-Math-7B/grpo/config_simple_rl.yaml b/recipes/Qwen2.5-Math-7B/grpo/config_simple_rl.yaml
@@ -37,8 +37,8 @@ num_generations: 7
 num_train_epochs: 1
 output_dir: data/Qwen-2.5-7B-Simple-RL
 overwrite_output_dir: true
-per_device_eval_batch_size: 2
-per_device_train_batch_size: 2
+per_device_eval_batch_size: 16
+per_device_train_batch_size: 16
 push_to_hub: true
 report_to:
 - wandb