docs: fix typo, formatting, and relative links in docs

Author: Kanaricc

README.md

@@ -38,7 +38,7 @@ We aim to build a easy-to-learn Agent tuner that unlock more possibilities for a
 - **Rich Tutorial Library**. AgentJet provides a rich library of [examples](https://github.com/modelscope/AgentJet/tree/main/tutorial) as tutorials.
 - **Efficient and Scalable**. AgentJet uses [verl] as the default backbone (`--backbone=verl`). However, we also support [trinity](https://github.com/modelscope/Trinity-RFT/) as alternative backbone, accelerating your tuning process via fully asynchronous RFT.
 - **Flexible and Fast**. AgentJet supports [multi-agent workflows](https://doc.agentjet.top/AgentJet/en/workflow.md) and adopts a context merging technique, accelerating training by 1.5x to 10x when the workflow involves multi-turn (or multi-agent) conversations.
-- **Reliability and Reproducibility**. Our team keeps track of framework performance across multiple [tasks + major-git-version + training-backbones](https://benchmark.agent-matrix.com/) (under construction, still gathering data, comming soon).
+- **Reliability and Reproducibility**. Our team keeps track of framework performance across multiple [tasks + major-git-version + training-backbones](https://benchmark.agent-matrix.com/) (under construction, still gathering data, coming soon).
 
 For advanced researchers, AgentJet also provides high-resolution logging and debugging solutions:
 <!-- For advanced researchers, AgentJet provides high-resolution logging and debugging solutions that are, to our knowledge, unprecedented in other prior projects. -->

docs/en/example_learning_to_ask.md

@@ -9,10 +9,8 @@ Train an agent to **ask the next best question** (instead of answering directly)
 
 In **Learning to Ask**, each training sample is a short **doctor–patient chat history**. The agent outputs **one next question** the doctor should ask next (optionally with multiple-choice answers), rather than giving diagnosis or treatment.
 
-```{figure} https://img.alicdn.com/imgextra/i4/O1CN01m9WJCM1WJL1aJCSaS_!!6000000002767-2-tps-1024-559.png
-Figure: Why "Learning to Ask" matters. Left: LLM gives a diagnosis with too little information. Right: LLM asks clear follow-up questions before concluding, which feels more reassuring.
-```
-
+![](https://img.alicdn.com/imgextra/i4/O1CN01m9WJCM1WJL1aJCSaS_!!6000000002767-2-tps-1024-559.png)
+<center><small>Why "Learning to Ask" matters. Left: LLM gives a diagnosis with too little information. Right: LLM asks clear follow-up questions before concluding, which feels more reassuring.</small></center>
 
 
 This tutorial is organized in two steps:

docs/en/example_werewolves.md

@@ -7,6 +7,7 @@ This tutorial demonstrates how to train **multiple agents** to play the Werewolv
 The Werewolves role-playing game is a typical POMDP (Partially Observable Markov Decision Process) problem. We can train agents in this cooperative multi-agent problem using shared-parameter methods.
 
 Terms explained:
+
 - **Partially Observable**: Agents are only able to receive **local information**. One agent cannot obtain others' perception, even if they are teammates.
 - **Markov Decision Process**: Making decisions according to current situations.
 - **Shared-parameter**: Using one model as policy for multiple agents. But notice agents **share** policy (model parameters) but **do not share** perception (model input).
@@ -18,6 +19,7 @@ Terms explained:
 This page shows how to use the Werewolves social deduction game as a multi-agent environment to prepare data and environment, write an AgentScope Workflow, configure the reward module (Judge), and complete the full process from local debugging to formal training.
 
 Scenario Overview
+
 - Scenario: Classic Werewolves game, including roles such as werewolf, villager, seer, witch, and hunter.
 - Goal: Train a specific role (in this example, the `werewolf`) to achieve a higher win rate in games.
 
@@ -70,6 +72,7 @@ When `--backbone=debug`, Ray is disabled. You can use a VSCode `.vscode/launch.j
 ### 3.1 Core Process
 
 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 by using the trainable model (via `tuner.as_agentscope_model(...)`), while opponents use the fixed model.
@@ -112,11 +115,13 @@ When `judge_protocol: null`, training relies on the reward (or win/loss outcome)
 In `ExampleWerewolves.execute()`, the workflow first runs a full game by calling `werewolves_game(players, roles)`, and obtains `good_guy_win` (whether the good-guy side wins).
 
 Then it uses a **turn-level sparse win/loss reward**:
+
 - If `good_guy_win == True` and the training target is not `werewolf` (i.e., you are training a good-guy role), then `raw_reward = 1` and `is_success = True`.
 - If `good_guy_win == False` and the training target is `werewolf` (i.e., you are training a werewolf-side role), then `raw_reward = 1` and `is_success = True`.
 - Otherwise, the training side did not win: `raw_reward = 0` and `is_success = False`.
 
 Exception / invalid-behavior penalty:
+
 - If an exception is thrown during the game (e.g., the game cannot proceed), all trainable targets are penalized uniformly: `raw_reward = -0.1` and `is_success = False`.
 
 If you need a more fine-grained evaluation (e.g., giving partial credit for key intermediate decisions instead of only win/loss), implement a custom Judge and enable it via `ajet.task_judge.judge_protocol`.

docs/en/workflow.md

@@ -80,10 +80,10 @@ Next, use the `tuner` argument, call its `tuner.as_agentscope_model()` method:
 ### 3. Code Example
 
 <div class="card-grid">
-<a href="en/example_math_agent/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/mdi:calculator-variant.svg" class="card-icon card-icon-math" alt=""><h3>Math Agent</h3></div><p class="card-desc">Training a math agent that can write Python code to solve mathematical problems.</p></a>
-<a href="en/example_learning_to_ask/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/mdi:comment-question.svg" class="card-icon card-icon-general" alt=""><h3>Learning to Ask</h3></div><p class="card-desc">Learning to ask questions like a doctor for medical consultation scenarios.</p></a>
-<a href="en/example_countdown/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/mdi:timer-sand.svg" class="card-icon card-icon-tool" alt=""><h3>Countdown Game</h3></div><p class="card-desc">Writing a countdown game using AgentScope and solving it with RL.</p></a>
-<a href="en/example_frozenlake/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/mdi:snowflake.svg" class="card-icon card-icon-data" alt=""><h3>Frozen Lake</h3></div><p class="card-desc">Solving a frozen lake walking puzzle using AgentJet's reinforcement learning.</p></a>
+<a href="../example_math_agent/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/mdi:calculator-variant.svg" class="card-icon card-icon-math" alt=""><h3>Math Agent</h3></div><p class="card-desc">Training a math agent that can write Python code to solve mathematical problems.</p></a>
+<a href="../example_learning_to_ask/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/mdi:comment-question.svg" class="card-icon card-icon-general" alt=""><h3>Learning to Ask</h3></div><p class="card-desc">Learning to ask questions like a doctor for medical consultation scenarios.</p></a>
+<a href="../example_countdown/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/mdi:timer-sand.svg" class="card-icon card-icon-tool" alt=""><h3>Countdown Game</h3></div><p class="card-desc">Writing a countdown game using AgentScope and solving it with RL.</p></a>
+<a href="../example_frozenlake/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/mdi:snowflake.svg" class="card-icon card-icon-data" alt=""><h3>Frozen Lake</h3></div><p class="card-desc">Solving a frozen lake walking puzzle using AgentJet's reinforcement learning.</p></a>
 </div>