verifiers 0.4.1.dev8__tar.gz → 0.4.1.dev12__tar.gz

{verifiers-0.4.1.dev8 → verifiers-0.4.1.dev12}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.4.1.dev8
+Version: 0.4.1.dev12
 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-0.4.1.dev8 → verifiers-0.4.1.dev12}/tests/integration/test_skills.py

@@ -2,20 +2,11 @@
 installs (skill-codeword-v1), across the harness x runtime grid."""
 
 import pytest
-from grid import grid, skip_harness
-
-# TODO: fix skill discovery in the upstream nano-rlm harness, then drop this skip.
-SKILLS = skip_harness(
-    grid(needs=("skills",)),
-    "rlm",
-    "nano-rlm often fails to discover an installed SKILL.md within the turn cap",
-)
+from grid import grid
 
 
 @pytest.mark.parametrize(
-    "harness,harness_runtime",
-    SKILLS,
-    indirect=True,
+    "harness,harness_runtime", grid(needs=("skills",)), indirect=True
 )
 async def test_skills(run, harness, harness_runtime):
     (episode,) = await run(

{verifiers-0.4.1.dev8 → verifiers-0.4.1.dev12}/tests/unit/test_trace.py

@@ -241,3 +241,28 @@ def test_transcript(make_full_trace):
     assert "[tool search]\nTOOL RESULT: Paris is the capital." in transcript
     assert "It is Paris." in transcript
     assert "SECRET REASONING" not in transcript  # reasoning is excluded
+
+
+def test_sampling_mask_from_packed_matches_lists():
+    import base64
+
+    import numpy as np
+
+    from verifiers.v1.types import SamplingMask
+
+    rows = [[4, 7], [], [9]]
+    ids, counts = np.array([4, 7, 9], np.int32), np.array([2, 0, 1], np.int32)
+    packed = {
+        name: {
+            "data": base64.b64encode(a.tobytes()).decode(),
+            "shape": list(a.shape),
+            "dtype": "int32",
+        }
+        for name, a in (("ids", ids), ("counts", counts))
+    }
+    for mask in (
+        SamplingMask.from_sampling_mask(rows),
+        SamplingMask.from_sampling_mask(packed),
+    ):
+        np.testing.assert_array_equal(mask.ids, ids)
+        np.testing.assert_array_equal(mask.counts, counts)

{verifiers-0.4.1.dev8 → verifiers-0.4.1.dev12}/verifiers/v1/harnesses/rlm/harness.py

@@ -31,7 +31,6 @@ BuiltinTool = Literal["bash", "edit", "fetch", "ipython"]
 
 RLM_REPO = "github.com/PrimeIntellect-ai/nano-rlm.git"
 RLM_CACHE_DIR = "/tmp/vf-rlm"
-SKILLS_DIR = "/task/rlm-skills"
 RLM_STATE_DIR = "/tmp/vf-rlm-state"
 RLM_RUNTIME_METADATA_KEY = "ai.prime.rlm/runtime-v1"
 RLM_SESSION_METADATA_KEY = "ai.prime.rlm/session-v1"
@@ -69,7 +68,7 @@ class CompactionConfig(BaseConfig):
 
 class RLMHarnessConfig(HarnessConfig):
     version: str = Field(
-        default="c38e20b60c90bbf625274d3fde2b2a384f390a66", min_length=1
+        default="77c3489d1a7f918b75c31e195480ed3556af60a5", min_length=1
     )
     """Git ref (branch, tag, or commit) of nano-rlm to install. Must know every
     field this harness puts on the wire, i.e. be at least the default ref."""
@@ -151,12 +150,14 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
         if self.config.skills:
             # Editable skill packages and uv's tool environment must belong to this run.
             self._skills_install_dir = f"{RLM_CACHE_DIR}-skills-{secrets.token_hex(16)}"
+            await self.install_skills(runtime, f"{self._skills_home()}/skills")
         directory = self._install_dir()
-        skills_dir = f"{directory}/skills" if self.config.skills else SKILLS_DIR
-        await self.install_skills(runtime, skills_dir)
         binary = f"{directory}/bin/rlm"
         checkout = f"{directory}/checkout"
         ready = f"{directory}/.ready"
+        # install.sh installs the Python skill packages under $RLM_HOME/skills.
+        home = self._skills_home()
+        home_env = f"RLM_HOME={home} " if home else ""
         # install.sh fetches curl/uv itself; add git only when the image lacks it.
         install = (
             f"rm -f {ready} && "
@@ -164,8 +165,7 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
             "{ apt-get update -qq && apt-get install -y -qq git; } && "
             f"rm -rf {checkout} && git clone https://{RLM_REPO} {checkout} && "
             f"git -C {checkout} checkout {shlex.quote(self.config.version)} && "
-            f"sed -i 's|/task/rlm-skills|{skills_dir}|g' {checkout}/install.sh && "
-            f"UV_INSTALL_DIR={directory}/bin UV_TOOL_BIN_DIR={directory}/bin "
+            f"{home_env}UV_INSTALL_DIR={directory}/bin UV_TOOL_BIN_DIR={directory}/bin "
             f"UV_TOOL_DIR={directory}/tools "
             f"RLM_CHECKOUT_PATH={checkout} bash {checkout}/install.sh && "
             f"touch {ready})"
@@ -296,9 +296,15 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
                 runtime, self._skills_install_dir, "RLM skill installation"
             )
 
-    @staticmethod
-    def _home(trace: Trace) -> str:
-        return f"{RLM_STATE_DIR}/{trace.id}/home"
+    def _home(self, trace: Trace) -> str:
+        return self._skills_home() or f"{RLM_STATE_DIR}/{trace.id}/home"
+
+    def _skills_home(self) -> str | None:
+        """rlm's home for a run that carries skills: they sit in its `skills/`, where
+        rlm installs and lists them, so the install and the process share it."""
+        if self._skills_install_dir is None:
+            return None
+        return f"{self._skills_install_dir}/home"
 
     def _install_dir(self) -> str:
         cache_key = hashlib.sha256(self.config.version.encode()).hexdigest()

{verifiers-0.4.1.dev8 → verifiers-0.4.1.dev12}/verifiers/v1/types.py

@@ -1,3 +1,4 @@
+import base64
 from collections.abc import Iterable
 from dataclasses import dataclass
 from typing import Annotated, Any, Literal
@@ -208,7 +209,16 @@ class SamplingMask:
     counts: Any
 
     @classmethod
-    def from_sampling_mask(cls, sampling_mask: list[list[int]]) -> "SamplingMask":
+    def from_sampling_mask(
+        cls, sampling_mask: list[list[int]] | dict[str, Any]
+    ) -> "SamplingMask":
+        """Build from vLLM's one-list-per-token masks, or from prime-rl's packed CSR
+        form: `{"ids", "counts"}`, each a base64 `{data, shape, dtype}` int32 array."""
+        if isinstance(sampling_mask, dict):
+            return cls(
+                ids=_decode_packed(sampling_mask["ids"], np.int32),
+                counts=_decode_packed(sampling_mask["counts"], np.int32),
+            )
         counts = np.fromiter(
             (len(row) for row in sampling_mask),
             dtype=np.int32,
@@ -222,6 +232,11 @@ class SamplingMask:
         return cls(ids=ids, counts=counts)
 
 
+def _decode_packed(packed: dict[str, Any], dtype: type) -> np.ndarray:
+    array = np.frombuffer(base64.b64decode(packed["data"]), dtype=packed["dtype"])
+    return array.astype(dtype, copy=False)
+
+
 class TurnTokens(BaseModel):
     """Training tokens from renderer tokenization or provider-returned token IDs."""