verifiers 0.1.15.dev388__tar.gz → 0.1.15.dev390__tar.gz

{verifiers-0.1.15.dev388 → verifiers-0.1.15.dev390}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.1.15.dev388
+Version: 0.1.15.dev390
 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-0.1.15.dev388 → verifiers-0.1.15.dev390}/verifiers/v1/ARCHITECTURE.md

@@ -68,10 +68,14 @@ several. A branch *is* a training sample: concatenating its nodes' `token_ids` r
 `logprobs` line up — no agent-specific export code. This is why compaction and subagents train
 end to end: each surviving context window is just another root→leaf path.
 
-`Trace.to_wire()` (`trace.py`) is `model_dump` minus what shouldn't travel: computed views
-(`reward`, `branches`, `num_turns`), per-span `duration`s (recomputed on the far side), and
-`usage`. Multimodal pixel tensors *do* travel — base64-encoded by a field serializer on
-`MessageNode.multi_modal_data` (`graph.py`) so they survive JSON + msgpack to the trainer.
+`Trace.to_record()` (`trace.py`) is the JSON record dump (`model_dump(mode="json")`) for
+`results.jsonl` / W&B tables, minus the per-node training tensors (`MessageNode.multi_modal_data`,
+`routed_experts`, via `_NODE_DUMP_EXCLUDE`): those hold raw numpy bytes that can't round-trip JSON
+(the dump raises `UnicodeDecodeError` on real expert ids) and bloat every line. Computed views
+(`reward`, `branches`, `num_turns`, per-span `duration`) are pydantic properties, so they're never
+serialized and recompute on load; `state` is excluded. The tensors still reach the trainer over the
+env-server *wire*, which uses msgpack `model_dump(mode="python")` and carries them as raw `bin` bytes
+(not base64) via the field serializers on `MessageNode` (`graph.py`); only the JSON record strips them.
 
 ### Branching: message-level vs renderer-level, and the token invariant
 

{verifiers-0.1.15.dev388 → verifiers-0.1.15.dev390}/verifiers/v1/harnesses/rlm/harness.py

@@ -11,6 +11,7 @@ programmatically (`await tools_<name>(...)`), rather than via a native MCP clien
 import json
 import logging
 import shlex
+from typing import Literal
 
 from pydantic import model_validator
 
@@ -22,6 +23,8 @@ from verifiers.v1.trace import Trace
 
 logger = logging.getLogger(__name__)
 
+BuiltinSkill = Literal["edit", "search"]
+
 RLM_REPO = "github.com/PrimeIntellect-ai/rlm.git"
 # rlm writes its session under $RLM_HOME/sessions/<id>/; point it at a workdir-
 # relative dir so it stays in the runtime (and is cleaned up with the workdir).
@@ -37,8 +40,8 @@ class RLMHarnessConfig(HarnessConfig):
     """Git ref (branch, tag, or commit) of rlm to install."""
     max_depth: int = 0
     """Recursion depth rlm may spawn sub-harnesses to (RLM_MAX_DEPTH)."""
-    skills: list[str] | None = None
-    """Built-in rlm skills to enable (RLM_SKILLS), e.g. `["edit"]`; None enables none.
+    skills: list[BuiltinSkill] = []
+    """Built-in rlm skills to enable (RLM_SKILLS), e.g. `["edit"]`; empty enables none.
     The tool set is fixed (ipython); only built-in skills are selectable."""
 
     @model_validator(mode="after")

{verifiers-0.1.15.dev388 → verifiers-0.1.15.dev390}/verifiers/v1/trace.py

@@ -201,6 +201,16 @@ class Branch(StrictBaseModel):
         return usage.completion_tokens if usage else 0
 
 
+_NODE_DUMP_EXCLUDE: dict = {
+    "nodes": {"__all__": {"multi_modal_data", "routed_experts"}}
+}
+"""The per-node fields `Trace.to_record` strips from a JSON dump: the multimodal `mm_kwargs`
+carrier and the router-replay `routed_experts` array. Both hold raw numpy bytes (msgpack `bin`)
+that the env-server wire carries fine but JSON cannot — `model_dump(mode="json")` raises
+`UnicodeDecodeError` on real (>0x7F) expert ids — and they bloat every record line besides.
+Lives with the schema it excludes so a new tensor field can't silently start crashing dumps."""
+
+
 class Trace(StrictBaseModel, Generic[TaskT, StateT]):
     """The full record of one rollout. Subclass to add typed fields."""
 
@@ -229,7 +239,7 @@ class Trace(StrictBaseModel, Generic[TaskT, StateT]):
     """Transient per-rollout runtime state (see `verifiers.v1.state.State`): shared with the tool/user
     servers as `self.state` (synced over the interception server) and read+written by scoring. Runtime
     scratch (counters, game progress, end-of-trajectory flags) — excluded from every dump (`model_dump`
-    + `to_wire`), unlike `info` which persists. Type it via `Taskset[Task, Config, MyState]`; defaults
+    / `to_record`), unlike `info` which persists. Type it via `Taskset[Task, Config, MyState]`; defaults
     to the base `State`."""
 
     is_completed: bool = False
@@ -401,6 +411,18 @@ class Trace(StrictBaseModel, Generic[TaskT, StateT]):
         )
         self.stop("error")
 
+    def to_record(self) -> dict[str, Any]:
+        """A JSON-serializable record of this rollout for `results.jsonl` / W&B tables.
+
+        `model_dump(mode="json")` minus the per-node training tensors (`_NODE_DUMP_EXCLUDE`):
+        those carry raw numpy bytes that JSON can't encode (the plain dump raises
+        `UnicodeDecodeError` on real expert ids) and that bloat every line. Everything else —
+        including `token_ids` / `mask` / `logprobs` / `is_content` and provider `usage` — is
+        kept; `state` and the computed-property views (`reward`, `branches`, `duration`) are
+        absent by construction (excluded / never serialized). The tensors still reach the
+        trainer over the env-server wire (msgpack raw bytes); only this record path strips them."""
+        return self.model_dump(mode="json", exclude=_NODE_DUMP_EXCLUDE)
+
 
 TraceT = TypeVar("TraceT", bound=Trace)  # type: ignore[type-arg]