diff --git a/training/configs/external_generalisation/math_curriculum_qwen_7b.yaml b/training/configs/external_generalisation/math_curriculum_qwen_7b.yaml
index 546a8d7b..b8291b3f 100644
--- a/training/configs/external_generalisation/math_curriculum_qwen_7b.yaml
+++ b/training/configs/external_generalisation/math_curriculum_qwen_7b.yaml
@@ -86,6 +86,7 @@ actor_rollout_ref:
     ppo_max_token_len_per_gpu: 49152 # n * ${data.max_prompt_length} + ${data.max_response_length}
     grad_clip: 1.0
     clip_ratio: 0.2
+    loss_agg_mode: "token-mean" # / "seq-mean-token-sum" / "seq-mean-token-mean"
     entropy_coeff: 0.001
     use_kl_loss: False # True for GRPO
     kl_loss_coef: 0.001 # for grpo