rnow 0.2.4__py3-none-any.whl

rnow/templates/mcp-tavily/config.yml

@@ -0,0 +1,29 @@
+project_id: ""
+project_name: "SimpleQA - Factual QA Agent"
+dataset_id: ""
+dataset_name: "simpleqa"
+dataset_type: rl
+organization_id: ""
+data:
+  train_file: train.jsonl
+  batch_size: 32
+  group_size: 16
+model:
+  path: Qwen/Qwen3-8B
+  qlora_rank: 32
+  name: "SimpleQA Agent"
+  description: "Multi-turn RL model trained on SimpleQA factual questions using Tavily MCP"
+algorithm:
+  loss_fn: ppo
+  adv_estimator: grpo
+  kl_penalty_coef: 0.01
+rollout:
+  max_turns: 3
+  max_tokens: 4096
+  termination_policy: last_tool
+  mcp_url: "https://mcp.tavily.com/mcp/?tavilyApiKey=YOUR_TAVILY_API_KEY"
+  max_tool_response_chars: 4000
+trainer:
+  num_epochs: 4
+  learning_rate: 0.0001
+  save_step: 63

rnow/templates/mcp-tavily/requirements.txt

@@ -0,0 +1 @@
+jellyfish>=1.0.0

rnow/templates/mcp-tavily/rewards.py

@@ -0,0 +1,25 @@
+import re
+
+import jellyfish
+
+from rnow.core import RewardArgs, reward
+
+
+@reward
+def accuracy(args: RewardArgs, messages: list) -> float:
+    """Check if the final answer matches the expected answer."""
+    response = messages[-1].get("content", "")
+    expected = args.metadata.get("expected_answer", "").strip().lower()
+
+    # Extract content after "Final Answer:"
+    match = re.search(r"Final Answer:\s*(.+?)(?:\n|$)", response, re.IGNORECASE)
+    if not match:
+        return 0.0
+
+    answer = match.group(1).strip().lower()
+
+    # Use Jaro-Winkler similarity (1.0 = exact match, 0.0 = no similarity)
+    similarity = jellyfish.jaro_winkler_similarity(answer, expected)
+
+    # Require high similarity (>0.9) to count as correct
+    return 1.0 if similarity > 0.9 else 0.0