diff --git a/training/configs/external_generalisation/math_curriculum_qwen_7b.yaml b/training/configs/external_generalisation/math_curriculum_qwen_7b.yaml
index 26cd4989..3fb47c6c 100644
--- a/training/configs/external_generalisation/math_curriculum_qwen_7b.yaml
+++ b/training/configs/external_generalisation/math_curriculum_qwen_7b.yaml
@@ -64,8 +64,8 @@ data:
   prompt_key: prompt
   max_prompt_length: 4096
   max_response_length: 2048
-  train_batch_size: 32
-  val_batch_size: 64
+  train_batch_size: 16
+  val_batch_size: 8
   return_raw_chat: True
   return_raw_input_ids: True
 
@@ -79,9 +79,9 @@ actor_rollout_ref:
     use_remove_padding: True
   actor:
     strategy: fsdp  # This is for backward-compatibility
-    ppo_mini_batch_size: 32
+    ppo_mini_batch_size: 4
     ppo_micro_batch_size: null # will be deprecated, use ppo_micro_batch_size_per_gpu
-    ppo_micro_batch_size_per_gpu: 8
+    ppo_micro_batch_size_per_gpu: 4
     use_dynamic_bsz: False
     ppo_max_token_len_per_gpu: 49152 # n * ${data.max_prompt_length} + ${data.max_response_length}
     grad_clip: 1.0