update docs

Author: binary-husky

README.md

@@ -67,7 +67,7 @@ uv pip install flash_attn==2.8.1 --no-build-isolation --no-cache-dir
 You can start training your first agent with a single command using a pre-configured YAML file. Take the [Math agent](docs/en/example_math_agent.md) as an example:
 
 ```bash
-ajet --conf tutorial/example_math_agent/math_agent.yaml --backbone='trinity' --with-ray
+ajet --conf tutorial/example_math_agent/math_agent.yaml
 ```
 
 #### Example Library

docs/en/example_app_world.md

@@ -43,7 +43,7 @@ export APPWORLD_SCRIPT="bash EnvService/env_sandbox/appworld.sh"
 Run the training script:
 
 ```bash
-ajet --conf tutorial/example_appworld/appworld.yaml --backbone='trinity' --with-ray
+ajet --conf tutorial/example_appworld/appworld.yaml --with-appworld
 ```
 
 <details>
@@ -90,13 +90,13 @@ This section explains how the AppWorld example is assembled: workflow, reward, c
 
 The AgentScope workflow code for the AppWorld example is located at `tutorial/example_appworld/appworld.py`.
 
-The code first defines the AgentScope workflow (set the agent's `model` to `model_tuner`):
+The code first defines the AgentScope workflow (set the agent's `model` to `tuner.as_agentscope_model()`):
 
 ```python
 agent = ReActAgent(
     name="Qwen",
     sys_prompt=first_msg["content"],
-    model=model_tuner,
+    model=tuner.as_agentscope_model(),
     formatter=DashScopeChatFormatter(),
     memory=InMemoryMemory(),
     toolkit=None,
@@ -105,7 +105,7 @@ agent = ReActAgent(
 
 env = workflow_task.gym_env
 
-for step in range(model_tuner.config.ajet.rollout.multi_turn.max_steps):
+for step in range(tuner.config.ajet.rollout.multi_turn.max_steps):
     # agentscope deal with interaction message
     reply_message = await agent(interaction_message)
     # env service protocol
@@ -117,14 +117,14 @@ for step in range(model_tuner.config.ajet.rollout.multi_turn.max_steps):
     # is terminated?
     if terminate:
         break
-    if model_tuner.get_context_tracker().context_overflow:
+    if tuner.get_context_tracker().context_overflow:
         break
 ```
 
 In the above code:
 
 - `env.step`: simulates the gym interface. It takes an action as input and returns a four-tuple `(observation, reward, terminate_flag, info)`.
-- `model_tuner.get_context_tracker().context_overflow`: checks whether the current context window has exceeded the token limit.
+- `tuner.get_context_tracker().context_overflow`: checks whether the current context window has exceeded the token limit.
 
 
 ### 3.2 Reward

docs/en/example_countdown.md

@@ -84,7 +84,7 @@ See details in `tutorial/example_countdown/countdown.py`. You can create new Age
 self.agent = ReActAgent(
     name="countdown_react_agent",
     sys_prompt=system_prompt,
-    model=model_tuner,
+    model=tuner.as_agentscope_model(),
     formatter=DashScopeChatFormatter(),
     memory=InMemoryMemory(),
     max_iters=2,

docs/en/example_frozenlake.md

@@ -23,13 +23,13 @@ pip install gymnasium[toy_text]
 Use the provided configuration file to quickly start training:
 
 ```bash
-ajet --conf tutorial/example_frozenlake/frozenlake_easy.yaml --backbone='trinity' --with-ray
+ajet --conf tutorial/example_frozenlake/frozenlake_easy.yaml --backbone='verl'
 ```
 
 To try a harder setting:
 
 ```bash
-ajet --conf tutorial/example_frozenlake/frozenlake_hard.yaml --backbone=trinity --with-ray
+ajet --conf tutorial/example_frozenlake/frozenlake_hard.yaml --backbone='verl'
 ```
 
 <details>

docs/en/example_learning_to_ask.md

@@ -42,6 +42,8 @@ After preprocessing, you should have: `train.jsonl` and`test.jsonl`。
 #### 2.2 Start Training
 
 ```bash
+ajet --conf tutorial/example_learn2ask/learn2ask.yaml --backbone='verl'
+# or
 ajet --conf tutorial/example_learn2ask/learn2ask.yaml --backbone='trinity' --with-ray
 ```
 

docs/en/example_math_agent.md

@@ -35,7 +35,7 @@ python scripts/download_dataset.py --target=openai/gsm8k --path=/the/path/to/sto
 # (optional) recommended cleanup before training
 # ajet --kill="python|ray|vllm"
 
-ajet --conf tutorial/example_math_agent/math_agent.yaml --backbone='trinity' --with-ray
+ajet --conf tutorial/example_math_agent/math_agent.yaml --backbone='verl'
 ```
 
 ??? tip "Quick Debugging (Optional)"
@@ -135,7 +135,7 @@ self.toolkit.register_tool_function(execute_python_code)
 self.agent = ReActAgent(
     name="math_react_agent",
     sys_prompt=system_prompt,
-    model=model_tuner,  # trainer-managed model wrapper
+    model=tuner.as_agentscope_model(),  # trainer-managed model wrapper
     formatter=DashScopeChatFormatter(),
     toolkit=self.toolkit,
     memory=InMemoryMemory(),

docs/en/example_tracing_feedback_loop.md

@@ -75,7 +75,8 @@ ajet:
 When everything is ready, start the training with `launcher.py`.
 
 ```bash
-# this launch the demo
+ajet --conf tutorial/example_feedback_tracing/example_feedback_tracing.yaml --backbone='verl'
+# or
 ajet --conf tutorial/example_feedback_tracing/example_feedback_tracing.yaml --backbone='trinity' --with-ray
 ```
 

docs/en/example_werewolves.md

@@ -26,7 +26,7 @@ Scenario Overview
 Start training with the following command:
 ```
 # ( ajet --kill="python|ray|vllm" )
-ajet --conf tutorial/example_werewolves/werewolves.yaml --backbone='trinity' --with-ray
+ajet --conf tutorial/example_werewolves/werewolves.yaml --backbone='verl'
 ```
 
 <details>
@@ -72,9 +72,9 @@ When `--backbone=debug`, Ray is disabled. You can use a VSCode `.vscode/launch.j
 At a high level, each training iteration follows this flow:
 - The task reader generates a new game setup (players, role assignments, initial state).
 - The rollout runs the AgentScope workflow to simulate a full game.
-- Agents in `trainable_targets` act using the trainable model (via `model_tuner`), while opponents use the fixed model.
+- Agents in `trainable_targets` act by using the trainable model (via `tuner.as_agentscope_model(...)`), while opponents use the fixed model.
 - The environment produces rewards / outcomes for the episode.
-- Trajectories are collected and passed to the backbone trainer (e.g., `trinity`) to update the trainable model.
+- Trajectories are collected and passed to the backbone trainer (`verl` or `trinity`) to update the trainable model.
 
 ### 3.2 Configuration Details
 

docs/en/intro.md

@@ -51,7 +51,7 @@ We recommend using `uv` for dependency management.
 You can start training your first agent with a single command using a pre-configured YAML file:
 
 ```bash
-ajet --conf tutorial/example_math_agent/math_agent.yaml --backbone='trinity' --with-ray
+ajet --conf tutorial/example_math_agent/math_agent.yaml
 ```
 
 !!! example "Learn More"

docs/en/introduction.md

@@ -51,7 +51,7 @@ uv pip install flash_attn==2.8.1 --no-build-isolation --no-cache-dir
 You can start training your first agent with a single command using a pre-configured YAML file. Take the [Math agent](./example_math_agent.md) as an example:
 
 ```bash
-ajet --conf tutorial/example_math_agent/math_agent.yaml --backbone='trinity' --with-ray
+ajet --conf tutorial/example_math_agent/math_agent.yaml
 ```
 
 #### Example Library