Simplify Buffer and Explorer Config

benchmark/config/countdown-template.yaml

@@ -35,11 +35,9 @@ buffer:
       rollout_args:
         temperature: 1.0
         logprobs: 0
+      default_workflow_type: math_workflow
+      default_reward_fn_type: countdown_reward
     eval_tasksets: []
-    default_workflow_type: math_workflow
-    default_reward_fn_type: countdown_reward
-    system_prompt: null
-    reply_prefix: null
   trainer_input:
     experience_buffer:
       name: experience_buffer

benchmark/config/gsm8k-template.yaml

@@ -40,11 +40,9 @@ buffer:
       rollout_args:
         temperature: 1.0
         logprobs: 0
+      default_workflow_type: math_workflow
+      default_reward_fn_type: math_reward
     eval_tasksets: []
-    default_workflow_type: math_workflow
-    default_reward_fn_type: math_reward
-    system_prompt: null
-    reply_prefix: null
   trainer_input:
     experience_buffer:
       name: experience_buffer
@@ -79,7 +77,7 @@ trainer:
   enable_preview: true
   grad_clip: 1.0
   use_dynamic_bsz: true
-  ppo_max_token_len_per_gpu: 10240
+  max_token_len_per_gpu: 10240
   ulysses_sequence_parallel_size: 1
 monitor:
   monitor_type: wandb

docs/sphinx_doc/source/tutorial/example_async_mode.md

@@ -39,14 +39,14 @@ buffer:
         response_key: 'answer'
       rollout_args:
         temperature: 1.0
-    default_workflow_type: 'math_workflow'
+      default_workflow_type: 'math_workflow'
   trainer_input:
     experience_buffer:
       name: gsm8k_buffer
       storage_type: queue
       path: 'sqlite:///gsm8k.db'
 explorer:
-  runner_num: 32
+  runner_per_model: 8
   rollout_model:
     engine_num: 4
 synchronizer:
@@ -86,7 +86,7 @@ buffer:
         response_key: 'answer'
       rollout_args:
         temperature: 1.0
-    default_workflow_type: 'math_workflow'
+      default_workflow_type: 'math_workflow'
   trainer_input:
     experience_buffer:
       name: gsm8k_buffer
@@ -98,7 +98,7 @@ synchronizer:
 trainer:
   grad_clip: 1.0
   use_dynamic_bsz: true
-  ppo_max_token_len_per_gpu: 16384
+  max_token_len_per_gpu: 16384
   ulysses_sequence_parallel_size: 1
 ```
 
@@ -133,7 +133,7 @@ cluster:  # important
   gpu_per_node: 8
 explorer:
   name: 'explorer_new'  # important
-  runner_num: 64
+  runner_per_model: 8
   rollout_model:
     engine_num: 8
 buffer:
@@ -150,7 +150,7 @@ buffer:
         response_key: 'answer'
       rollout_args:
         temperature: 1.0
-    default_workflow_type: 'math_workflow'
+      default_workflow_type: 'math_workflow'
   trainer_input:
     experience_buffer:
       name: gsm8k_buffer

docs/sphinx_doc/source/tutorial/example_reasoning_basic.md

@@ -77,6 +77,7 @@ buffer:
         response_key: 'answer'
       rollout_args:
         temperature: 1.0
+      default_workflow_type: 'math_workflow'
     eval_tasksets:
     - name: gsm8k-eval
       storage_type: file
@@ -86,15 +87,15 @@ buffer:
       format:
         prompt_key: 'question'
         response_key: 'answer'
-    default_workflow_type: 'math_workflow'
+      default_workflow_type: 'math_workflow'
   trainer_input:
     experience_buffer:
       name: gsm8k_buffer
       storage_type: queue
       path: 'sqlite:///gsm8k.db'
 explorer:
   eval_interval: 50
-  runner_num: 16
+  runner_per_model: 16
   rollout_model:
     engine_num: 1
 synchronizer:
@@ -117,7 +118,7 @@ trinity run --config examples/grpo_gsm8k/gsm8k.yaml
 
 ## Optional: RFT with SFT Warmup
 
-Before RFT, we may use SFT as a warmup step. Trinity-RFT supports adding SFT warmup stage before RFT by setting `stages` in the config file. The `sft_warmup_dataset` specifies the dataset used for SFT warmup, and `sft_warmup_steps` specifies the number of training steps for SFT warmup.
+Before RFT, we may use SFT as a warmup step. Trinity-RFT supports adding SFT warmup stage before RFT by setting `stages` in the config file. The `experience_buffer` specifies the dataset used for SFT warmup, and `total_steps` specifies the number of training steps for SFT warmup.
 
 ```yaml
 # Properly add the following configs in gsm8k.yaml

docs/sphinx_doc/source/tutorial/example_step_wise.md

@@ -121,15 +121,15 @@ buffer:
       workflow_args:
         max_env_steps: 30
       enable_progress_bar: false
-    default_workflow_type: 'step_wise_alfworld_workflow'
+      default_workflow_type: 'step_wise_alfworld_workflow'
   trainer_input:
     experience_buffer:
       name: alfworld_buffer
       storage_type: queue
       use_priority_queue: true
 explorer:
   max_repeat_times_per_runner: 1
-  runner_num: 32
+  runner_per_model: 32
   max_timeout: 3600
   rollout_model:
     enable_history: true
@@ -152,7 +152,7 @@ trainer:
   save_interval: 50
   grad_clip: 1.0
   use_dynamic_bsz: true
-  ppo_max_token_len_per_gpu: 16384
+  max_token_len_per_gpu: 16384
   ulysses_sequence_parallel_size: 1
 ```
 

docs/sphinx_doc/source/tutorial/trinity_configs.md

@@ -200,6 +200,7 @@ buffer:
   batch_size: 32
   train_batch_size: 256
   total_epochs: 100
+  total_steps: null
 
   explorer_input:
     taskset:
@@ -214,9 +215,6 @@ buffer:
         ...
       buffer_2:
         ...
-
-  default_workflow_type: 'math_workflow'
-  default_reward_fn_type: 'countdown_reward'
 ```
 
 - `batch_size`: Number of tasks used per training step. *Please do not multiply this value by the `algorithm.repeat_times` manually*.
@@ -231,6 +229,9 @@ Defines the dataset(s) used by the explorer for training and evaluation.
 ```yaml
 buffer:
   explorer_input:
+    default_workflow_type: 'math_workflow'
+    default_eval_workflow_type: 'math_workflow'
+    default_reward_fn_type: 'countdown_reward'
     taskset:
       name: countdown_train
       storage_type: file
@@ -262,7 +263,10 @@ buffer:
 ```
 
 - `buffer.explorer_input.taskset`: Task dataset used for training exploration policies.
-- `buffer.explorer_input.eval_taskset`: List of task datasets used for evaluation.
+- `buffer.explorer_input.eval_tasksets`: List of task datasets used for evaluation.
+- `buffer.explorer_input.default_workflow_type`: Default workflow type for all task datasets under `explorer_input` if not specified at the dataset level.
+- `buffer.explorer_input.default_eval_workflow_type`: Default evaluation workflow type for all eval task datasets under `explorer_input` if not specified at the dataset level.
+- `buffer.explorer_input.default_reward_fn_type`: Default reward function type for all task datasets under `explorer_input` if not specified at the dataset level.
 
 The configuration for each task dataset is defined as follows:
 
@@ -413,7 +417,7 @@ trainer:
   save_strategy: "unrestricted"
   grad_clip: 1.0
   use_dynamic_bsz: true
-  ppo_max_token_len_per_gpu: 16384
+  max_token_len_per_gpu: 16384
   ulysses_sequence_parallel_size: 1
   trainer_config: null
 ```
@@ -429,7 +433,7 @@ trainer:
   - `unrestricted`: No restrictions on saving operations; multiple nodes, processes, or threads are allowed to save the model simultaneously.
 - `grad_clip`: Gradient clipping for updates.
 - `use_dynamic_bsz`: Whether to use dynamic batch size.
-- `ppo_max_token_len_per_gpu`:  The maximum number of tokens to be processed in forward and backward when updating the policy. Effective when `use_dynamic_bsz=true`.
+- `max_token_len_per_gpu`:  The maximum number of tokens to be processed in forward and backward when updating the policy. Effective when `use_dynamic_bsz=true`.
 - `ulysses_sequence_parallel_size`: Sequence parallel size.
 - `trainer_config`: The trainer configuration provided inline.
 ---

docs/sphinx_doc/source_zh/tutorial/example_async_mode.md

@@ -39,14 +39,14 @@ buffer:
         response_key: 'answer'
       rollout_args:
         temperature: 1.0
-    default_workflow_type: 'math_workflow'
+      default_workflow_type: 'math_workflow'
   trainer_input:
     experience_buffer:
       name: gsm8k_buffer
       storage_type: queue
       path: 'sqlite:///gsm8k.db'
 explorer:
-  runner_num: 32
+  runner_per_model: 16
   rollout_model:
     engine_num: 4
 synchronizer:
@@ -86,7 +86,7 @@ buffer:
         response_key: 'answer'
       rollout_args:
         temperature: 1.0
-    default_workflow_type: 'math_workflow'
+      default_workflow_type: 'math_workflow'
   trainer_input:
     experience_buffer:
       name: gsm8k_buffer
@@ -133,7 +133,7 @@ cluster:  # important
   gpu_per_node: 8
 explorer:
   name: 'explorer_new'  # important
-  runner_num: 64
+  runner_per_model: 8
   rollout_model:
     engine_num: 8
 buffer:
@@ -150,7 +150,7 @@ buffer:
         response_key: 'answer'
       rollout_args:
         temperature: 1.0
-    default_workflow_type: 'math_workflow'
+      default_workflow_type: 'math_workflow'
   trainer_input:
     experience_buffer:
       name: gsm8k_buffer

docs/sphinx_doc/source_zh/tutorial/example_reasoning_basic.md

@@ -77,6 +77,7 @@ buffer:
         response_key: 'answer'
       rollout_args:
         temperature: 1.0
+      default_workflow_type: 'math_workflow'
     eval_tasksets:
     - name: gsm8k-eval
       storage_type: file
@@ -86,15 +87,15 @@ buffer:
       format:
         prompt_key: 'question'
         response_key: 'answer'
-    default_workflow_type: 'math_workflow'
+      default_workflow_type: 'math_workflow'
   trainer_input:
     experience_buffer:
       name: gsm8k_buffer
       storage_type: queue
       path: 'sqlite:///gsm8k.db'
 explorer:
   eval_interval: 50
-  runner_num: 16
+  runner_per_model: 16
   rollout_model:
     engine_num: 1
 synchronizer:
@@ -117,7 +118,7 @@ trinity run --config examples/grpo_gsm8k/gsm8k.yaml
 
 ## 进阶选项：带 SFT warmup 的 RFT
 
-在进行 RFT 之前，我们可以先使用 SFT 作为预热步骤。Trinity-RFT 支持通过在配置文件中设置 `stages` 来添加 SFT 预热阶段。`sft_warmup_dataset` 指定用于 SFT warmup 的数据集，`sft_warmup_steps` 指定 SFT warmup 的训练步数。
+在进行 RFT 之前，我们可以先使用 SFT 作为预热步骤。Trinity-RFT 支持通过在配置文件中设置 `stages` 来添加 SFT 预热阶段。`experience_buffer` 指定用于 SFT warmup 的数据集，`total_steps` 指定 SFT warmup 的训练步数。
 
 ```yaml
 # 在 gsm8k.yaml 中正确添加以下配置

docs/sphinx_doc/source_zh/tutorial/example_step_wise.md

@@ -119,15 +119,15 @@ buffer:
       workflow_args:
         max_env_steps: 30
       enable_progress_bar: false
-    default_workflow_type: 'step_wise_alfworld_workflow'
+      default_workflow_type: 'step_wise_alfworld_workflow'
   trainer_input:
     experience_buffer:
       name: alfworld_buffer
       storage_type: queue
       use_priority_queue: true
 explorer:
   max_repeat_times_per_runner: 1
-  runner_num: 32
+  runner_per_model: 16
   max_timeout: 3600
   rollout_model:
     enable_history: true
@@ -150,7 +150,7 @@ trainer:
   save_interval: 50
   grad_clip: 1.0
   use_dynamic_bsz: true
-  ppo_max_token_len_per_gpu: 16384
+  max_token_len_per_gpu: 16384
   ulysses_sequence_parallel_size: 1
 ```
 

docs/sphinx_doc/source_zh/tutorial/trinity_configs.md

@@ -214,9 +214,6 @@ buffer:
         ...
       buffer_2:
         ...
-
-  default_workflow_type: 'math_workflow'
-  default_reward_fn_type: 'countdown_reward'
 ```
 
 - `batch_size`: 每个训练步骤使用的任务数。*请勿手动将此值乘以 `algorithm.repeat_times`*。
@@ -231,6 +228,9 @@ buffer:
 ```yaml
 buffer:
   explorer_input:
+    default_workflow_type: 'math_workflow'
+    default_eval_workflow_type: 'math_workflow'
+    default_reward_fn_type: 'countdown_reward'
     taskset:
       name: countdown_train
       storage_type: file
@@ -256,13 +256,14 @@ buffer:
         response_key: 'answer'
       rollout_args:
         temperature: 0.1
-      default_workflow_type: 'math_workflow'
-      default_reward_fn_type: 'countdown_reward'
     ...
 ```
 
 - `buffer.explorer_input.taskset`: 用于训练探索策略的任务数据集。
-- `buffer.explorer_input.eval_taskset`: 用于评估的任务数据集列表。
+- `buffer.explorer_input.eval_tasksets`: 用于评测的任务数据集列表。
+- `buffer.explorer_input.default_workflow_type`: 若未在数据集级别指定，则为所有任务数据集设置默认的工作流类型。
+- `buffer.explorer_input.default_eval_workflow_type`: 若未在数据集级别指定，则为所有评测任务数据集设置默认的工作流类型。
+- `buffer.explorer_input.default_reward_fn_type`: 若未在数据集级别指定，则为所有任务数据集设置默认的奖励类型。
 
 每个任务数据集的配置定义如下：
 
@@ -413,7 +414,7 @@ trainer:
   save_strategy: "unrestricted"
   grad_clip: 1.0
   use_dynamic_bsz: true
-  ppo_max_token_len_per_gpu: 16384
+  max_token_len_per_gpu: 16384
   ulysses_sequence_parallel_size: 1
   trainer_config: null
 ```
@@ -429,7 +430,7 @@ trainer:
   - `unrestricted`：不限制保存操作，允许多个节点、进程或线程同时保存模型。
 - `grad_clip`: 梯度裁剪阈值。
 - `use_dynamic_bsz`: 是否使用动态批量大小。
-- `ppo_max_token_len_per_gpu`: 训练过程中，每个 GPU 最大 token 长度; 当 `use_dynamic_bsz=true` 时生效。
+- `max_token_len_per_gpu`: 训练过程中，每个 GPU 最大 token 长度; 当 `use_dynamic_bsz=true` 时生效。
 - `ulysses_sequence_parallel_size`: 序列并行的并行度，即用于分割单个序列的 GPU 数量。
 - `trainer_config`: 内联提供的 trainer 配置。
 

examples/RAFT_alfworld/RAFT_alfworld_7B.yaml

@@ -70,7 +70,7 @@ trainer:
   save_interval: 100000
   grad_clip: 1.0
   use_dynamic_bsz: true
-  ppo_max_token_len_per_gpu: 20000  # Adjusted for alfworld longer sequences
+  max_token_len_per_gpu: 20000  # Adjusted for alfworld longer sequences
   ulysses_sequence_parallel_size: 1
 monitor:
   monitor_type: wandb

examples/RAFT_alfworld/RAFT_reflect_alfworld_7B.yaml

@@ -70,7 +70,7 @@ trainer:
   save_interval: 100000
   grad_clip: 1.0
   use_dynamic_bsz: true
-  ppo_max_token_len_per_gpu: 20000  # Adjusted for alfworld longer sequences
+  max_token_len_per_gpu: 20000  # Adjusted for alfworld longer sequences
   ulysses_sequence_parallel_size: 1
 monitor:
   monitor_type: wandb