rnow 0.2.4__py3-none-any.whl

rnow/templates/rl-tools/config.yml

@@ -0,0 +1,27 @@
+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 with tool use"
+algorithm:
+  loss_fn: ppo
+  adv_estimator: grpo
+  kl_penalty_coef: 0.01
+rollout:
+  max_turns: 3
+  max_tokens: 4096
+  termination_policy: last_tool
+trainer:
+  num_epochs: 4
+  learning_rate: 0.0001
+  save_step: 63

rnow/templates/rl-tools/env.py

@@ -0,0 +1,38 @@
+import requests
+from bs4 import BeautifulSoup
+
+from rnow.core.tool import tool
+
+
+@tool
+def internet_search(query: str) -> dict:
+    """Search the web and return up to 5 results (title, link, snippet)."""
+    try:
+        resp = requests.get(
+            "https://en.wikipedia.org/w/api.php",
+            params={
+                "action": "query",
+                "list": "search",
+                "srsearch": query,
+                "format": "json",
+                "srlimit": 5,
+            },
+            headers={"User-Agent": "ReinforceNow/1.0 (training platform)"},
+            timeout=10,
+        )
+        resp.raise_for_status()
+    except requests.RequestException:
+        return []
+    data = resp.json()
+    results = []
+    for item in data.get("query", {}).get("search", []):
+        snippet = BeautifulSoup(item.get("snippet", ""), "html.parser").get_text()
+        title = item.get("title", "")
+        results.append(
+            {
+                "title": title,
+                "link": f"https://en.wikipedia.org/wiki/{title.replace(' ', '_')}",
+                "snippet": snippet[:200],
+            }
+        )
+    return results

rnow/templates/rl-tools/requirements.txt

@@ -0,0 +1,3 @@
+requests>=2.28.0
+beautifulsoup4>=4.12.0
+jellyfish>=1.0.0

rnow/templates/rl-tools/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