Add examples for frozenlake and emailsearch (#94)

tuner/email_search/config.yaml

@@ -0,0 +1,72 @@
+project: "AgentScope" # Project name
+name: "Email_search" # Experiment name
+checkpoint_root_dir: ${oc.env:TRINITY_CHECKPOINT_ROOT_DIR,./checkpoints}  # Directory to save model checkpoints
+algorithm:
+  algorithm_type: multi_step_grpo  # GRPO series for multi-step scenario
+  repeat_times: 8  # Number of rollouts per prompt for advantage estimation
+  optimizer:
+    lr: 1e-6  # Learning rate
+  policy_loss_fn: "rec"  # Policy loss function
+  policy_loss_fn_args: # Policy loss function arguments
+    epsilon_low: 0.2
+    epsilon_high: 0.2
+    clip_mode: "one-side"
+    weight: "none"
+    temp: 1.0
+    regularizer: "none"
+    regularizer_coef: 0.0
+  kl_loss_fn: 'k2' # KL divergence loss function
+  kl_loss_fn_args:
+    kl_coef: 0.0 # KL divergence coefficient
+  advantage_fn_args:
+    std_cal_level: 'batch' # Advantage normalization level
+model:
+  model_path: ${oc.env:TRINITY_MODEL_PATH,Qwen/Qwen3-4B-Instruct-2507}  # Base model path
+  max_response_tokens: 4096  # Max tokens per response
+  max_model_len: 20480  # Max context length
+buffer:
+  total_epochs: 10  # Total training epochs
+  batch_size: 64  # Batch size per explore step
+  train_batch_size: 2560  # 64*8*5, total experiences per training step
+  trainer_input:
+    experience_buffer:
+      name: experience_buffer
+      storage_type: queue
+      replay_buffer:
+        enable: true  # Enable experience replay
+        priority_fn: 'decay_limit_randomization'
+        priority_fn_args:
+          decay: 2.0
+          use_count_limit: 3
+          sigma: 2.0
+explorer:
+  eval_interval: 10
+  max_repeat_times_per_runner: 1 # Max repeat times per runner
+  max_timeout: 3600 # Max timeout for each rollout (seconds)
+  rollout_model:
+    enable_history: true # Enable conversation history
+    enable_openai_api: true # Enable OpenAI-compatible API
+    enable_auto_tool_choice: true # Enable automatic tool selection
+    tool_call_parser: hermes # Parser for tool calls
+    engine_num: 4   # Number of vLLM engines for rollout model
+    tensor_parallel_size: 1  # TP size per engine for rollout model
+    enable_prefix_caching: false # Disable prefix caching
+  auxiliary_models:
+    - name: judge
+      model_path: Qwen/Qwen3-30B-A3B-Instruct-2507 # Judge model path
+      engine_num: 1  # Number of vLLM engines for judge model
+      tensor_parallel_size: 2  # TP size per engine for judge model
+      enable_thinking: false # Disable thinking/reasoning mode
+      max_prompt_tokens: 2048 # Max tokens for prompt
+      max_response_tokens: 128 # Max tokens for response
+      max_model_len: 2500 # Max model context length
+synchronizer:
+  sync_style: dynamic_by_explorer # Sync triggered dynamically by explorer
+  sync_interval: 5 # Sync every N steps
+  sync_timeout: 3600 # Timeout for synchronization (seconds)
+trainer:
+  save_interval: 100 # Save checkpoint every N steps
+  grad_clip: 1.0 # Gradient clipping value
+  use_dynamic_bsz: true # Use dynamic batch size
+  max_token_len_per_gpu: 16384 # Max token length per GPU
+  ulysses_sequence_parallel_size: 1 # Sequence parallel size for Ulysses