verifiers 0.2.2.dev67__tar.gz → 0.2.2.dev68__tar.gz

{verifiers-0.2.2.dev67 → verifiers-0.2.2.dev68}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.2.2.dev67
+Version: 0.2.2.dev68
 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-0.2.2.dev67 → verifiers-0.2.2.dev68}/tests/v1/test_e2e.py

@@ -104,6 +104,7 @@ async def test_single_turn(run_v1, harness, harness_runtime, tmp_path):
         "echo-v1",
         harness=harness,
         runtime={"type": harness_runtime},
+        env={"agent": {"sampling": {"temperature": 0.0}}},
         output_dir=tmp_path,
         max_turns=2,
     )
@@ -112,7 +113,9 @@ async def test_single_turn(run_v1, harness, harness_runtime, tmp_path):
     assert trace.stop_condition == "agent_completed"
     assert trace.reward == 1.0
     # The seat's resolved identity rides the trace (policy metadata for trainers).
-    assert trace.agent is not None and trace.agent.config.sampling.max_tokens == 2048
+    assert trace.agent is not None
+    assert trace.agent.config.sampling.max_tokens == 2048
+    assert trace.agent.config.sampling.temperature == 0.0
     # Every sampled turn has one per-call record, linked to its assistant node.
     sampled = [i for i, n in enumerate(trace.nodes) if n.sampled]
     assert [c.node for c in trace.calls if c.error is None] == sampled

{verifiers-0.2.2.dev67 → verifiers-0.2.2.dev68}/verifiers/v1/configs/agent.py

@@ -35,7 +35,7 @@ class AgentConfig(BaseConfig):
     client: ClientConfig | None = None
     """Endpoint override (None = the run's client)."""
     sampling: SamplingConfig | None = None
-    """Sampling override (None = the run's sampling)."""
+    """Sampling values merged onto the run's sampling."""
 
     max_turns: int | None = None
     """Max model turns per run (None = no limit)."""

{verifiers-0.2.2.dev67 → verifiers-0.2.2.dev68}/verifiers/v1/env.py

@@ -32,7 +32,7 @@ from verifiers.v1.mcp import SharedToolServer, serve_shared
 from verifiers.v1.runtimes import SubprocessConfig, runtime_is_local
 from verifiers.v1.task import Task
 from verifiers.v1.trace import Error, Trace
-from verifiers.v1.utils.generic import concrete_type
+from verifiers.v1.utils.generic import concrete_type, deep_merge
 from verifiers.v1.utils.memory import trim_memory_periodically
 from verifiers.v1.utils.retries import run_episode_with_retry
 
@@ -204,15 +204,21 @@ class Env(ABC, Generic[ConfigT]):
 
         def make(name: str, spec: AgentConfig) -> Agent:
             # Unpinned fields fall back to the run's ctx / the taskset's harness.
+            sampling = ctx.sampling
+            if spec.sampling is not None:
+                sampling = sampling.model_copy(
+                    update=deep_merge(
+                        sampling.model_dump(exclude_unset=True),
+                        spec.sampling.model_dump(exclude_unset=True),
+                    )
+                )
             resolved = spec.model_copy(
                 update={
                     "harness": spec.harness
                     if spec.harness is not None
                     else self._default_harness,
                     "model": spec.model if spec.model is not None else ctx.model,
-                    "sampling": spec.sampling
-                    if spec.sampling is not None
-                    else ctx.sampling,
+                    "sampling": sampling,
                 }
             )
             return _EpisodeAgent(