doc fix

Author: chenyushuo

examples/grpo_frozen_lake/frozen_lake.yaml

@@ -43,8 +43,8 @@ buffer:
           env_max_steps: 8
           agent_max_steps: 10
           is_slippery: false
+        repeat_times: 4
         rollout_args:
-          n: 4
           top_p: 0.8
           top_k: 20
     default_workflow_type: 'frozen_lake_workflow'

trinity/common/config.py

@@ -84,7 +84,7 @@ class GenerationConfig:
     logprobs: Optional[int] = None  # 0  # vLLM return `logprobs + 1` elements
     max_tokens: Optional[int] = None  # if None, use model.max_response_tokens
     # repeat each task for `n` times
-    # ! DO NOT SET, it will be set by `algorithm.repeat_times` or `max(buffer.explorer_input.eval_tasksets[i].repeat_times)`
+    # ! DO NOT SET, it will be set by `algorithm.repeat_times` or `buffer.explorer_input.eval_tasksets[i].repeat_times`
     n: int = 1
 
 
@@ -249,10 +249,10 @@ class TasksetConfig:
 
     enable_progress_bar: bool = False
 
+    # ! This setting is only valid for `eval_taskset`; for other taskset, it will be overridden by `algorithm.repeat_times`.
+    repeat_times: int = 1
     # ! DO NOT SET, automatically load from checkpoint
     index: int = 0
-    # ! DO NOT SET in trainer_input, automatically set from `algorithm.repeat_times`
-    repeat_times: int = 1
     # ! DO NOT SET, automatically set based on train/eval
     is_eval: bool = False
     # ! DO NOT SET, automatically set from buffer.batch_size