stop rollouts on incomplete responses (no content or tools)

.gitignore

@@ -33,6 +33,9 @@ __pycache__/
 
 # libraries
 /prime-rl
+/packages/tasksets
+/packages/harnesses
+
 
 # outputs
 wandb/

docs/environments.md

@@ -567,7 +567,7 @@ class MyGameEnv(vf.MultiTurnEnv):
         return state.get("lives", 1) <= 0
 ```
 
-`MultiTurnEnv` includes built-in stop conditions for errors, prompt length limits, and `max_turns` by default.
+`MultiTurnEnv` includes built-in stop conditions for errors, prompt length limits, `max_turns`, and incomplete response detection by default.
 
 Execution order can be controlled with `priority` (higher runs first). This is useful for checking cheap conditions before expensive ones:
 

environments/wiki_search/pyproject.toml

@@ -3,9 +3,9 @@ name = "wiki-search"
 description = "Agentic RAG over Wikipedia pages for trivia Q&A"
 tags = ["wikipedia", "multi-turn", "agentic-search", "rag", "train", "eval", "llm-judge"]
 requires-python = ">=3.11"
-version = "0.1.23"
+version = "0.1.24"
 dependencies = [
-    "verifiers>=0.1.9",
+    "verifiers>=0.1.11.dev0",
     "chromadb",
     "datasets",
     "openai",

environments/wiki_search/wiki_search.py

@@ -260,7 +260,10 @@ async def read_section(section_id: str) -> str:
     )
 
     async def judge_reward_func(judge, prompt, completion, answer, state) -> float:
-        judge_response = await judge(prompt, completion, answer, state)
+        cleaned_completion = [
+            {x["role"]: x["content"].split("</think>")[-1] for x in completion}
+        ]
+        judge_response = await judge(prompt, cleaned_completion, answer, state)
         if "yes" in judge_response.lower():
             return 1.0
         else:

verifiers/envs/environment.py

@@ -156,7 +156,9 @@ def __init__(
 
         if dataset is not None:
             if callable(dataset):
-                self.dataset_source: DatasetBuilder | None = dataset
+                self.dataset_source: DatasetBuilder | None = cast(
+                    DatasetBuilder, dataset
+                )
             else:
                 self.dataset_source = lambda ds=dataset: ds
                 self.build_dataset()  # Eagerly build for raw datasets (backwards compat)
@@ -165,7 +167,9 @@ def __init__(
 
         if eval_dataset is not None:
             if callable(eval_dataset):
-                self.eval_dataset_source: DatasetBuilder | None = eval_dataset
+                self.eval_dataset_source: DatasetBuilder | None = cast(
+                    DatasetBuilder, eval_dataset
+                )
             else:
                 self.eval_dataset_source = lambda ds=eval_dataset: ds
                 self.build_eval_dataset()  # Eagerly build for raw datasets (backwards compat)

verifiers/envs/multiturn_env.py

@@ -64,6 +64,10 @@ async def has_final_env_response(self, state: State) -> bool:
         """Check if env_response signaled termination via final_env_response."""
         return state.get("final_env_response") is not None
 
+    @vf.stop
+    async def has_incomplete_response(self, state: State) -> bool:
+        return state.get("incomplete_response", False)
+
     async def setup_state(self, state: State) -> State:
         """Override to add environment-specific state fields."""
         return state
@@ -121,9 +125,15 @@ async def add_model_response(
     ):
         completion_messages = await parse_response_message(response)
         tokens = await parse_response_tokens(response, self.max_seq_len)
+        has_content = bool(response.message.content)
+        has_tool_calls = bool(response.message.tool_calls)
+        if not has_content and not has_tool_calls:
+            state["incomplete_response"] = True
         response_is_truncated = response.message.is_truncated or False
-        is_truncated = response_is_truncated or (
-            tokens is not None and bool(tokens.get("is_truncated"))
+        is_truncated = (
+            response_is_truncated
+            or (tokens is not None and bool(tokens.get("is_truncated")))
+            or state.get("incomplete_response", False)
         )
         trajectory_step = TrajectoryStep(
             prompt=prompt_messages,