Add env and verifiers version to GenerateMetadata (#597)

docs/source/api_reference.md

@@ -60,7 +60,7 @@ class ProcessedOutputs(BaseModel):
     rewards: list[float]
 ```
 
-`GenerateOutputs.metadata` captures run-level context (environment + arguments, model + sampling configuration, summary statistics, and the resolved save path) so downstream tooling can reproduce or resume evaluations without guessing defaults.
+`GenerateOutputs.metadata` captures run-level context (environment + arguments, environment + library versions, model + sampling configuration, summary statistics, and the resolved save path) so downstream tooling can reproduce or resume evaluations without guessing defaults.
 
 ### State Dictionary
 
@@ -287,4 +287,3 @@ def reset_for_rollout(self, prompt: Messages, answer: str, info: Info | None) ->
     }
     return state
 ```
-

docs/source/training.md

@@ -228,4 +228,4 @@ However, the best way to improve training is ensuring appropriate task difficult
 
 - Explore [Environments](environments.md) to create custom tasks
 - Review [Components](components.md) for advanced patterns
-- See the [examples directory](https://github.com/PrimeIntellect-ai/verifiers/tree/main/examples) on GitHub for complete training scripts
+- See the [examples directory](https://github.com/PrimeIntellect-ai/verifiers/tree/main/examples) on GitHub for complete training scripts

tests/test_environment.py

@@ -7,7 +7,7 @@
 from datasets import Dataset
 from openai.types.chat.chat_completion import Choice
 
-from verifiers import Environment, Parser, Rubric, ThinkParser
+from verifiers import Environment, Parser, Rubric, ThinkParser, __version__ as vf_version
 from verifiers.types import (
     GenerateMetadata,
     GenerateOutputs,
@@ -76,6 +76,8 @@ def _make_metadata(
     return GenerateMetadata(
         env_id="test-env",
         env_args={},
+        env_version="test-env-version",
+        verifiers_version="test-verifiers-version",
         model="test-model",
         base_url="http://localhost",
         num_examples=num_examples,
@@ -420,6 +422,7 @@ def reward_b(**kwargs):
                 weights=[0.5, 0.5],
             ),
         )
+        env.version = "test-env-v1"
 
         results = await env.generate(
             inputs=env.get_dataset(n=2),
@@ -434,6 +437,8 @@ def reward_b(**kwargs):
         assert "reward_b" in results["metadata"]["avg_metrics"]
         assert results["metadata"]["avg_metrics"]["reward_a"] == 1.0
         assert results["metadata"]["avg_metrics"]["reward_b"] == 0.5
+        assert results["metadata"]["env_version"] == "test-env-v1"
+        assert results["metadata"]["verifiers_version"] == vf_version
 
     @pytest.mark.asyncio
     async def test_generate_metadata_without_scoring(self, mock_openai_client):

tests/test_environment_extra.py

@@ -88,6 +88,8 @@ def _make_metadata(
     return GenerateMetadata(
         env_id="dummy-env",
         env_args={},
+        env_version="dummy-env-version",
+        verifiers_version="test-verifiers-version",
         model="test-model",
         base_url="http://localhost",
         num_examples=num_examples,

tests/test_eval_cli.py

@@ -10,6 +10,8 @@ def _make_metadata(config) -> GenerateMetadata:
     return GenerateMetadata(
         env_id=config.env_id,
         env_args=config.env_args,
+        env_version="test-env-version",
+        verifiers_version="test-verifiers-version",
         model=config.model,
         base_url=config.client_config.api_base_url,
         num_examples=config.num_examples,

tests/test_eval_utils.py

@@ -15,6 +15,8 @@ def _make_metadata(
     return GenerateMetadata(
         env_id="test-env",
         env_args={},
+        env_version="test-env-version",
+        verifiers_version="test-verifiers-version",
         model="test-model",
         base_url="http://localhost",
         num_examples=num_examples,

verifiers/envs/environment.py

@@ -53,6 +53,38 @@
     pass
 
 
+def _get_verifiers_version() -> str | None:
+    try:
+        return pkg_version("verifiers")
+    except PackageNotFoundError:
+        try:
+            from verifiers import __version__
+
+            return __version__
+        except Exception:
+            return None
+
+
+def _get_env_version(env: "Environment") -> str | None:
+    # Instance override
+    if getattr(env, "version", None):
+        return str(getattr(env, "version"))
+
+    env_module = inspect.getmodule(env.__class__)
+    if env_module and hasattr(env_module, "__version__"):
+        version = getattr(env_module, "__version__")
+        if version:
+            return str(version)
+
+    top_level_module = env.__class__.__module__.split(".")[0]
+    try:
+        return pkg_version(top_level_module)
+    except PackageNotFoundError:
+        pass
+
+    return None
+
+
 class Environment(ABC):
     """
     Base class for all environments.
@@ -600,6 +632,8 @@ def _prepare_rollout_results(
         metadata = GenerateMetadata(
             env_id=self.env_id,
             env_args=self.env_args,
+            env_version=_get_env_version(self),
+            verifiers_version=_get_verifiers_version(),
             model=model,
             base_url=str(client.base_url) if hasattr(client, "base_url") else "",
             num_examples=num_unique_examples,

verifiers/types.py

@@ -142,6 +142,8 @@ class GenerateMetadata(TypedDict):
 
     env_id: str
     env_args: dict
+    env_version: str | None
+    verifiers_version: str | None
     model: str
     base_url: str
     num_examples: int