verifiers 0.2.2.dev51__tar.gz → 0.2.2.dev52__tar.gz

{verifiers-0.2.2.dev51 → verifiers-0.2.2.dev52}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.2.2.dev51
+Version: 0.2.2.dev52
 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.dev51 → verifiers-0.2.2.dev52}/tests/v1/conftest.py

@@ -135,10 +135,6 @@ def _eval_config(
     `{id: ...}` config; `runtime` places the `agent` seat's harness (an agent field, not a
     harness one); `model` overrides the default text model (e.g. a VLM for an image task).
 
-    `temperature=0` (greedy) makes the run reproducible; `max_tokens` is generous headroom,
-    not a target — these trivial tasks finish in a few hundred tokens, so capping tighter only
-    risks truncating the reasoning before the answer (which tanks the reward).
-
     `harness=None` leaves every seat on its own story — the multi-agent case: there
     is no run-level harness, so a single-agent test's `harness` lands on the `agent`
     seat and a multi-agent test pins its seats through `env` role fields instead."""
@@ -175,7 +171,6 @@ def _eval_config(
         num_rollouts=n,
         sampling={
             "max_tokens": max_tokens,
-            "temperature": 0,
             "reasoning_effort": reasoning_effort,
         },
         rich=False,
@@ -219,8 +214,8 @@ def run_v1_server():
 
 @pytest.fixture
 async def live_ctx():
-    """A live `ModelContext` (the e2e default model + endpoint, greedy) for driving
-    `Agent` directly — the agent-surface counterpart of `run_v1`."""
+    """A live `ModelContext` (the e2e default model + endpoint, provider-default
+    sampling) for driving `Agent` directly — the agent-surface counterpart of `run_v1`."""
     from verifiers.v1.clients import EvalClientConfig, ModelContext, resolve_client
     from verifiers.v1.types import SamplingConfig
 
@@ -229,7 +224,7 @@ async def live_ctx():
         yield ModelContext(
             model="deepseek/deepseek-v4-flash",
             client=client,
-            sampling=SamplingConfig(max_tokens=2048, temperature=0),
+            sampling=SamplingConfig(max_tokens=2048),
         )
     finally:
         await client.close()
@@ -252,7 +247,7 @@ def run_v0():
             legacy={"id": env_id, "args": args or {}},
             num_tasks=1,
             num_rollouts=n,
-            sampling={"max_tokens": max_tokens, "temperature": 0},
+            sampling={"max_tokens": max_tokens},
             rich=False,
             output_dir=output_dir,
         )

{verifiers-0.2.2.dev51 → verifiers-0.2.2.dev52}/tests/v1/test_e2e.py

@@ -109,7 +109,7 @@ 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.temperature == 0
+    assert trace.agent is not None and trace.agent.config.sampling.max_tokens == 2048
     # 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