verifiers 0.1.15.dev405__tar.gz → 0.1.15.dev406__tar.gz

{verifiers-0.1.15.dev405 → verifiers-0.1.15.dev406}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.1.15.dev405
+Version: 0.1.15.dev406
 Summary: Verifiers: Environments for LLM Reinforcement Learning
 Project-URL: Homepage, https://github.com/primeintellect-ai/verifiers
 Project-URL: Documentation, https://github.com/primeintellect-ai/verifiers

{verifiers-0.1.15.dev405 → verifiers-0.1.15.dev406}/verifiers/v1/GUIDE.md

@@ -66,7 +66,7 @@ class ReverseTaskset(vf.Taskset[ReverseTask, ReverseConfig]):
 
     @vf.reward()
     async def exact_match(self, task: ReverseTask, trace: vf.Trace) -> float:
-        return float(trace.assistant_messages[-1].content.strip() == task.answer)
+        return float(trace.last_reply == task.answer)
 
 
 __all__ = ["ReverseTaskset"]   # vf resolves the taskset by finding this Taskset subclass
@@ -250,7 +250,7 @@ VERIFY = (Path(__file__).parent / "verify.py").read_text()   # PEP 723 header de
 
 @vf.reward()
 async def verify(self, task, trace, runtime) -> float:
-    r = await runtime.run_uv_script(VERIFY, args=[task.answer, trace.assistant_messages[-1].content])
+    r = await runtime.run_uv_script(VERIFY, args=[task.answer, trace.last_reply])
     return float(r.stdout.strip() == "1.0")
 ```
 
@@ -317,7 +317,7 @@ a taskset `@vf.stop` fires. A stop is an `async (self, trace) -> bool` checked b
 ```python
 @vf.stop
 async def saw_answer(self, trace) -> bool:
-    last = trace.assistant_messages[-1].content or ""
+    last = trace.last_reply
     return "FINAL:" in last
 ```
 

{verifiers-0.1.15.dev405 → verifiers-0.1.15.dev406}/verifiers/v1/trace.py

@@ -364,6 +364,15 @@ class Trace(StrictBaseModel, Generic[TaskT, StateT]):
             if n.sampled and isinstance(n.message, AssistantMessage)
         ]
 
+    @property
+    def last_reply(self) -> str:
+        """The model's final reply as plain text — the last assistant message's
+        `content` (whitespace-stripped), or ``""`` when there are no assistant
+        messages or the last one has no text. Convenience over
+        ``(assistant_messages[-1].content or "").strip()`` for last-turn scoring."""
+        msgs = self.assistant_messages
+        return (msgs[-1].content or "").strip() if msgs else ""
+
     @property
     def tool_messages(self) -> list[ToolMessage]:
         """The tool results in the latest full context — the main (last) branch's