verifiers 0.3.2.dev42__tar.gz → 0.3.2.dev44__tar.gz

{verifiers-0.3.2.dev42 → verifiers-0.3.2.dev44}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev42
+Version: 0.3.2.dev44
 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-0.3.2.dev42 → verifiers-0.3.2.dev44}/pyproject.toml

@@ -107,33 +107,6 @@ required-version = ">=0.11.1"
 exclude-newer = "7 days"
 default-groups = ["dev", "examples"]
 
-# NeMo Gym declares its full platform stack; this integration only starts resource servers.
-[[tool.uv.dependency-metadata]]
-name = "nemo-gym"
-version = "0.4.0"
-requires-python = ">=3.12"
-requires-dist = [
-    "aiohttp>=3.14.1",
-    "fastapi",
-    "gprof2dot",
-    "hydra-core",
-    "itsdangerous",
-    "mcp>=1.27,<2",
-    "omegaconf",
-    "openai>=2.9",
-    "orjson",
-    "pandas",
-    "pydantic",
-    "pydot",
-    "ray>=2.55.1",
-    "requests",
-    "rich",
-    "typing-extensions",
-    "uvicorn",
-    "wandb",
-    "yappi",
-]
-
 [[tool.uv.index]]
 name = "pypi"
 url = "https://pypi.org/simple"
@@ -244,8 +217,6 @@ addopts = [
 ]
 markers = [
     "e2e: marks end-to-end eval-run tests (need a model API key; skipped without one)",
-    "integration: marks tests as integration tests",
-    "slow: marks slow-running tests (deselect with -m 'not slow')",
     # v1 e2e matrix axes — select subsets with `-m` (see tests/v1/conftest.py). A mark is applied
     # per axis, so it matches every case touching that value on any axis.
     "subprocess: v1 e2e cases touching the subprocess (host) runtime",
@@ -264,11 +235,7 @@ markers = [
     "pi: v1 e2e cases on the pi harness",
     "pool: v1 e2e cases on the pool harness",
     "codex: v1 e2e cases on the codex harness",
-    "unit: marks tests as unit tests",
     "asyncio: marks tests as async tests",
-    "parsers: marks tests for parser components",
-    "rubrics: marks tests for rubric components",
-    "environments: marks tests for environment components",
 ]
 filterwarnings = [
     "ignore::DeprecationWarning",
@@ -276,7 +243,6 @@ filterwarnings = [
     "ignore::UserWarning:transformers.*",
     "ignore:The 'repr' attribute.*:UserWarning:pydantic.*",
     "ignore:The 'frozen' attribute.*:UserWarning:pydantic.*",
-    "ignore:.*is.*with.*str.*literal:SyntaxWarning:cellpylib.*",
 ]
 asyncio_mode = "auto"
 norecursedirs = [".git", ".tox", "dist", "build", "*.egg", "__pycache__"]

{verifiers-0.3.2.dev42 → verifiers-0.3.2.dev44}/tests/v1/test_configs.py

@@ -1,9 +1,7 @@
 """Every checked-in v1 eval config parses.
 
 Mirrors prime-rl's config test: glob the configs and assert each validates into its config
-type. The root `configs/*.toml` are the `uv run eval @ <file>` v1 configs (EvalConfig);
-`endpoints.toml` isn't an eval config, and `configs/eval|rl|gepa/` are the legacy
-`vf-eval` / training formats (different, non-v1 config classes), so both are out of scope here.
+type. The root `configs/*.toml` are the `uv run eval @ <file>` v1 configs (EvalConfig).
 """
 
 import tomllib
@@ -13,11 +11,7 @@ import pytest
 
 from verifiers.v1.configs.cli.eval import EvalConfig
 
-CONFIGS = sorted(
-    p
-    for p in (Path(__file__).resolve().parents[2] / "configs").glob("*.toml")
-    if p.name != "endpoints.toml"
-)
+CONFIGS = sorted((Path(__file__).resolve().parents[2] / "configs").glob("*.toml"))
 
 
 @pytest.mark.parametrize("path", CONFIGS, ids=lambda p: p.name)

{verifiers-0.3.2.dev42 → verifiers-0.3.2.dev44}/verifiers/v1/agent.py

@@ -9,7 +9,7 @@ server; un-entered, each run brings its own."""
 
 import asyncio
 import logging
-from collections.abc import AsyncIterator, Callable, Iterator, Mapping
+from collections.abc import AsyncIterator, Callable, Mapping
 from contextlib import asynccontextmanager, nullcontext
 from dataclasses import dataclass
 from typing import Self
@@ -733,9 +733,3 @@ class Agents:
                 f"{sorted(agents) if agents else []}"
             )
         return agents[name]
-
-    def __iter__(self) -> Iterator[Agent]:
-        return iter(self._agents.values())
-
-    def __len__(self) -> int:
-        return len(self._agents)

{verifiers-0.3.2.dev42 → verifiers-0.3.2.dev44}/verifiers/v1/cli/eval/main.py

@@ -46,11 +46,9 @@ def main(argv: list[str] | None = None) -> None:
                 narrow_config(EvalConfig, argv)
             )  # full option help, narrowed to the given ids
         return
-    # An env-block flag (or a since-moved flat axis) skips the usage gate so the
-    # typed parse renders its did-you-mean instead of a bare usage line.
-    typed_axis = any(
-        a.startswith(("--env.", "--taskset.", "--harness.", "--serve.")) for a in argv
-    )
+    # An env-block flag skips the usage gate so the typed parse renders its
+    # did-you-mean instead of a bare usage line.
+    typed_axis = any(a.startswith(("--env.", "--serve.")) for a in argv)
     if (
         not extract_id(argv, "env.taskset")
         and not references_config_file(argv)

{verifiers-0.3.2.dev42 → verifiers-0.3.2.dev44}/verifiers/v1/cli/gepa.py

@@ -42,7 +42,7 @@ def main(argv: list[str] | None = None) -> None:
                 narrow_config(GEPAConfig, argv)
             )  # full option help, narrowed to the given ids
         return
-    typed_axis = any(a.startswith(("--env.", "--taskset.", "--harness.")) for a in argv)
+    typed_axis = any(a.startswith("--env.") for a in argv)
     if (
         not extract_id(argv, "env.taskset")
         and not references_config_file(argv)

{verifiers-0.3.2.dev42 → verifiers-0.3.2.dev44}/verifiers/v1/cli/output.py

@@ -72,13 +72,10 @@ def attempt_log_file(run_dir: Path) -> Path:
 
 
 def saved_config_path(run_dir: Path) -> Path | None:
-    """The run's saved resolved config (`configs/resolved/<cli>.json`; legacy runs
-    kept it at `configs/<cli>.json`), None if absent."""
-    for config_dir in (run_dir / CONFIG_DIR / RESOLVED_DIR, run_dir / CONFIG_DIR):
-        candidates = sorted(config_dir.glob("*.json")) if config_dir.is_dir() else []
-        if candidates:
-            return candidates[0]
-    return None
+    """The run's saved resolved config (`configs/resolved/<cli>.json`), None if absent."""
+    config_dir = run_dir / CONFIG_DIR / RESOLVED_DIR
+    candidates = sorted(config_dir.glob("*.json")) if config_dir.is_dir() else []
+    return candidates[0] if candidates else None
 
 
 # Compiling an adapter is the expensive part; run output reuses only a few model classes.

{verifiers-0.3.2.dev42 → verifiers-0.3.2.dev44}/verifiers/v1/episode.py

@@ -6,6 +6,7 @@ from typing import Annotated, Any, Generic, Literal, Self
 from pydantic import BaseModel, Field, model_validator
 
 from verifiers.v1.configs.agent import WireAgentConfig
+from verifiers.v1.graph import RECORD_FLOAT_DECIMALS
 from verifiers.v1.state import State, StateT
 from verifiers.v1.task import DataT, WireTaskData
 from verifiers.v1.trace import EXCLUDE_FIELDS, AgentConfigT, Error, Trace, TraceTask
@@ -146,12 +147,16 @@ class Episode(BaseModel, Generic[DataT, StateT, AgentConfigT]):
             grouped.setdefault(trace.agent.name, []).append(trace)
         return grouped
 
-    def to_record(self) -> dict[str, Any]:
-        """JSON record without raw trace tensors, which remain on the msgpack wire."""
+    def to_record(
+        self, float_decimals: int | None = RECORD_FLOAT_DECIMALS
+    ) -> dict[str, Any]:
+        """JSON record without raw trace tensors, which remain on the msgpack wire.
+        Per-token float streams are rounded to `float_decimals` (`None` keeps every digit)."""
         return self.model_dump(
             mode="json",
             exclude={"traces": {"__all__": EXCLUDE_FIELDS}},
             exclude_none=True,
+            context={"float_decimals": float_decimals},
         )
 
 

{verifiers-0.3.2.dev42 → verifiers-0.3.2.dev44}/verifiers/v1/graph.py

@@ -28,7 +28,14 @@ from dataclasses import dataclass
 from typing import TYPE_CHECKING, Any
 
 import numpy as np
-from pydantic import BaseModel, ConfigDict, Field, field_serializer, field_validator
+from pydantic import (
+    BaseModel,
+    ConfigDict,
+    Field,
+    FieldSerializationInfo,
+    field_serializer,
+    field_validator,
+)
 from pydantic.json_schema import SkipJsonSchema
 from renderers.base import MultiModalData, PlaceholderRange, RenderedTokens
 
@@ -65,6 +72,14 @@ def _decode_ndarray(d: dict) -> np.ndarray:
     return np.frombuffer(d["data"], dtype=np.dtype(d["dtype"])).reshape(d["shape"])
 
 
+RECORD_FLOAT_DECIMALS = 4
+"""Default precision of per-token float streams in JSON records (`to_record`). Full-precision
+digits are noise to every record reader and the least compressible bytes of a trace; four
+decimals leave a logprob within 1e-4 of what the trainer saw. `to_record(float_decimals=...)`
+overrides it per dump (`None` keeps every digit). The msgpack wire (`mode="python"`) keeps
+full precision — training never reads the record."""
+
+
 class MessageNode(BaseModel):
     """One message in the graph: a message plus the tokens it adds to the cumulative
     sequence. Concatenating a root→leaf path's nodes reconstructs that branch's full token
@@ -150,6 +165,22 @@ class MessageNode(BaseModel):
 
     model_config = ConfigDict(arbitrary_types_allowed=True)
 
+    @field_serializer(
+        "logprobs",
+        "advantages",
+        "reference_logprobs",
+        "trainer_logprobs",
+        "entropies",
+        when_used="json",
+    )
+    def serialize_record_floats(
+        self, values: list[float] | None, info: FieldSerializationInfo
+    ) -> list[float] | None:
+        decimals = (info.context or {}).get("float_decimals", RECORD_FLOAT_DECIMALS)
+        if values is None or decimals is None:
+            return values
+        return [round(value, decimals) for value in values]
+
     @field_serializer("multi_modal_data")
     def serialize_multi_modal_data(self, mmd: MultiModalData | None) -> dict | None:
         """`MultiModalData` -> msgpack-safe dict so the pixel tensors ride the wire; numpy

{verifiers-0.3.2.dev42 → verifiers-0.3.2.dev44}/verifiers/v1/harness.py

@@ -117,31 +117,6 @@ class Harness(ABC, Generic[ConfigT]):
                 # `write` moves bytes, not modes; restore the execute bits scripts need.
                 await runtime.run(["chmod", "+x", *executables], {})
 
-    async def run(
-        self,
-        ctx: ModelContext,
-        trace: Trace,
-        runtime: Runtime,
-        endpoint: str,
-        secret: str,
-        mcp_urls: dict[str, str],
-        data: TaskData,
-        messages: Messages | None = None,
-    ) -> None:
-        """Run ONE segment of the exchange without retaining a process: the program
-        from launch (or, with `messages`, the user's next turn(s) via `resume`) until
-        it yields. The rollout loop owns the exchange across segments."""
-        async with boundary(HarnessError, f"harness {self.config.id!r}"):
-            if messages is None:
-                result = await self.launch(
-                    ctx, trace, runtime, endpoint, secret, mcp_urls, data
-                )
-            else:
-                result = await self.resume(
-                    ctx, trace, runtime, endpoint, secret, mcp_urls, data, messages
-                )
-        await self._check_result(trace, runtime, result)
-
     async def _check_result(
         self, trace: Trace, runtime: Runtime, result: ProgramResult
     ) -> None:

{verifiers-0.3.2.dev42 → verifiers-0.3.2.dev44}/verifiers/v1/harnesses/prime_agent/harness.py

@@ -97,7 +97,6 @@ class PrimeAgentHarnessConfig(HarnessConfig):
 class PrimeAgentHarness(ACPHarness[PrimeAgentHarnessConfig]):
     APPENDS_SYSTEM_PROMPT = True
     SUPPORTS_MCP = True
-    SUPPORTS_RESUME = True
     SUPPORTS_SKILLS = True
 
     def acp_turn_result(self, trace: Trace, result: ACPTurn) -> None:

{verifiers-0.3.2.dev42 → verifiers-0.3.2.dev44}/verifiers/v1/mcp/launch.py

@@ -201,9 +201,8 @@ async def _install_in_sandbox(server: ServerBase, runtime: Runtime) -> str:
         await runtime.write(env_remote, env_data)
     venv = str(PurePosixPath(workdir) / ".vf-venv")
     root_q, temp_q, cache_q, venv_q = map(shlex.quote, (root, temp, cache, venv))
-    extras = ",".join(type(server).EXTRAS)
     vf_source = shlex.quote(vf_remote)
-    env_source = shlex.quote(env_remote + (f"[{extras}]" if extras else ""))
+    env_source = shlex.quote(env_remote)
     setup = (
         f"set -e; mkdir -p {root_q} {temp_q} {cache_q}; "
         f"export TMPDIR={temp_q} UV_CACHE_DIR={cache_q}; "
@@ -279,12 +278,7 @@ async def serve_in_runtime(
         env["MCP_PORT_FILE"] = port_file
     python = sys.executable
     if runtime.type != "subprocess":
-        # Prebuilt runtime images may already contain the server and its Python
-        # environment. Let those servers bypass redundant per-launch source
-        # uploads and installation while preserving the existing default.
-        python = getattr(type(server), "RUNTIME_PYTHON", None)
-        if python is None:
-            python = await _install_in_sandbox(server, runtime)
+        python = await _install_in_sandbox(server, runtime)
     command = [python, "-m", type(server).__module__]
     if runtime.type != "subprocess":
         # Providers may invoke uv after the install shell exits, so preserve its PATH.

{verifiers-0.3.2.dev42 → verifiers-0.3.2.dev44}/verifiers/v1/mcp/server.py

@@ -134,9 +134,6 @@ class ServerBase(Generic[ConfigT, StateT]):
     tools bare (no `<server>_` prefix); name collisions across servers are then the taskset
     author's concern."""
 
-    EXTRAS: ClassVar[tuple[str, ...]] = ()
-    """Package extras the server's module needs, applied at sandbox install."""
-
     def __init__(self, config: ConfigT) -> None:
         self.config = config
         self._state_cls = state_cls(type(self))

{verifiers-0.3.2.dev42 → verifiers-0.3.2.dev44}/verifiers/v1/trace.py

@@ -18,7 +18,7 @@ if TYPE_CHECKING:
 from verifiers.v1 import graph
 from verifiers.v1.configs.agent import AgentConfig, WireAgentConfig
 from verifiers.v1.errors import ProviderError
-from verifiers.v1.graph import MessageNode
+from verifiers.v1.graph import RECORD_FLOAT_DECIMALS, MessageNode
 from verifiers.v1.runtimes import RuntimeInfo
 from verifiers.v1.semantic import (
     ACPInfo,
@@ -728,9 +728,16 @@ class Trace(BaseModel, Generic[DataT, StateT, AgentConfigT]):
         self.ok = False
         self.stop("error")
 
-    def to_record(self) -> dict[str, Any]:
-        """JSON record without raw tensors, which remain available on the msgpack wire."""
-        return self.model_dump(mode="json", exclude=EXCLUDE_FIELDS)
+    def to_record(
+        self, float_decimals: int | None = RECORD_FLOAT_DECIMALS
+    ) -> dict[str, Any]:
+        """JSON record without raw tensors, which remain available on the msgpack wire.
+        Per-token float streams are rounded to `float_decimals` (`None` keeps every digit)."""
+        return self.model_dump(
+            mode="json",
+            exclude=EXCLUDE_FIELDS,
+            context={"float_decimals": float_decimals},
+        )
 
 
 WireTrace = Trace[WireTaskData, State, WireAgentConfig]