synth-ai 0.2.14__py3-none-any.whl → 0.4.1__py3-none-any.whl

synth_ai/sdk/task/contracts.py

@@ -0,0 +1,298 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from enum import Enum
+from typing import Any, Literal
+
+from pydantic import BaseModel, ConfigDict, Field
+
+
+class RolloutMode(str, Enum):
+    """Mode controls how rollout infrastructure processes inference URLs."""
+    RL = "rl"
+    EVAL = "eval"
+
+
+class OutputMode(str, Enum):
+    """Controls how the policy expects model outputs.
+
+    - TOOL_CALLS: Use function/tool calling (default, current behavior)
+    - TEXT: Plain text in message.content
+    - STRUCTURED: JSON via response_format (OpenAI json_schema, Groq json_object, Gemini responseSchema)
+    """
+    TOOL_CALLS = "tool_calls"
+    TEXT = "text"
+    STRUCTURED = "structured"
+
+
+class StructuredOutputConfig(BaseModel):
+    """Configuration for structured output mode (OutputMode.STRUCTURED).
+
+    Defines the JSON schema that the model must conform to when using structured outputs.
+    This is normalized across providers:
+    - OpenAI: response_format.json_schema
+    - Groq: response_format.json_schema or json_object
+    - Gemini: generationConfig.responseSchema
+    """
+    schema: dict[str, Any] = Field(
+        ...,
+        description="JSON Schema for the expected response structure"
+    )
+    schema_name: str = Field(
+        default="response",
+        description="Name for the schema (required by some providers)"
+    )
+    strict: bool = Field(
+        default=True,
+        description="Whether to enforce strict schema validation (OpenAI strict mode)"
+    )
+
+
+@dataclass(frozen=True)
+class TaskAppEndpoints:
+    """Required Task App endpoints used by RL trainers and clients.
+
+    Task Apps run as lightweight HTTP services (often on Modal) that expose these
+    standard endpoints. Additional endpoints (proxies, debug routes) may be added
+    by individual task apps as needed.
+    """
+
+    root: str = "/"
+    health: str = "/health"
+    info: str = "/info"
+    task_info: str = "/task_info"
+    rollout: str = "/rollout"
+
+
+# --- Unified rollout schema used by Task App services and SDK utilities ---
+
+
+class RolloutEnvSpec(BaseModel):
+    env_id: str | None = None
+    env_name: str | None = None
+    config: dict[str, Any] = Field(default_factory=dict)
+    seed: int | None = None
+
+
+class RolloutPolicySpec(BaseModel):
+    policy_id: str | None = None
+    policy_name: str | None = None
+    config: dict[str, Any] = Field(default_factory=dict)
+
+    # Output mode configuration (defaults to tool_calls for backward compatibility)
+    output_mode: OutputMode = Field(
+        default=OutputMode.TOOL_CALLS,
+        description="How the policy expects model outputs: tool_calls, text, or structured"
+    )
+    structured_config: StructuredOutputConfig | None = Field(
+        default=None,
+        description="Configuration for structured output mode (required if output_mode=STRUCTURED)"
+    )
+
+
+class RolloutRecordConfig(BaseModel):
+    trajectories: bool = True
+    logprobs: bool = False
+    value: bool = False
+    return_trace: bool = False
+    trace_format: Literal["compact", "full", "structured"] = "compact"
+
+
+class RolloutSafetyConfig(BaseModel):
+    max_ops: int = 100000
+    max_time_s: float = 3600.0
+
+
+class RolloutRequest(BaseModel):
+    run_id: str
+    env: RolloutEnvSpec
+    policy: RolloutPolicySpec
+    ops: list[dict[str, Any]] | list[str]
+    record: RolloutRecordConfig = RolloutRecordConfig()
+    on_done: str = "reset"
+    safety: RolloutSafetyConfig = RolloutSafetyConfig()
+    training_session_id: str | None = None
+    synth_base_url: str | None = None
+    mode: RolloutMode  # Required: explicit RL vs EVAL mode
+
+
+class RolloutStep(BaseModel):
+    """Single step in a rollout trajectory.
+
+    DEPRECATED: This is part of the legacy trajectory format. New code should
+    consume v3 traces (RolloutResponse.trace) instead. See monorepo/trace_single_source.txt
+    for migration plan.
+    """
+    obs: dict[str, Any]
+    tool_calls: list[dict[str, Any]] = Field(default_factory=list)
+    reward: float | None = None
+    done: bool = False
+    truncated: bool | None = None
+    info: dict[str, Any] | None = None
+
+    # Unified output fields (supports all output modes)
+    output: dict[str, Any] | str | None = Field(
+        default=None,
+        description="Unified output: parsed JSON for STRUCTURED, raw text for TEXT, tool args for TOOL_CALLS"
+    )
+    output_mode: OutputMode | None = Field(
+        default=None,
+        description="Which output mode produced this step's output"
+    )
+
+
+class RolloutTrajectory(BaseModel):
+    """Legacy trajectory format for rollout results.
+    
+    DEPRECATED: This format duplicates data already present in v3 traces and will
+    be removed once training code migrates to consuming RolloutResponse.trace.
+    
+    Current state:
+    - Task apps emit BOTH this format AND v3 traces (dual serialization)
+    - Training code (GSPO) reads from this format
+    - Eval/filter tools read from v3 traces
+    
+    Migration plan:
+    - Phase 1: Training code learns to read from v3 traces (with fallback to this)
+    - Phase 2: Make this field optional once training is migrated
+    - Phase 3: Remove this field entirely and delete this class
+    
+    See: monorepo/trace_single_source.txt for full migration plan and timeline.
+    
+    Why v3 traces are better:
+    - Single source of truth (no duplication/drift)
+    - Richer data: token IDs, logprobs, reasoning, timing, images
+    - Built-in audit trail and replay capability
+    - Standard schema across all Synth AI tooling
+    """
+    env_id: str
+    policy_id: str
+    steps: list[RolloutStep]
+    final: dict[str, Any] | None = None
+    length: int
+    
+    # Required for trace correlation with inference mesh (optional initially for backward compat)
+    # See: monorepo/INFERENCE_URL_REQUIREMENT_PLAN.md and trace_creation_and_judgement.txt
+    inference_url: str
+
+    # Required by monorepo trace_validation.py: trajectory-level trace with event_history
+    # The event_history contains LM call records for input/output extraction
+    trace: dict[str, Any] | None = Field(
+        default=None,
+        description="V3 trace with event_history for this trajectory (required for trace strict mode)"
+    )
+
+    decision_samples: list[dict[str, Any]] | None = None
+
+
+class RolloutMetrics(BaseModel):
+    episode_returns: list[float]
+    mean_return: float
+    num_steps: int
+    num_episodes: int = 0
+    outcome_score: float | None = None
+    events_score: float | None = None
+    details: dict[str, Any] = Field(default_factory=dict)
+
+
+class RolloutResponse(BaseModel):
+    """Response from a rollout execution.
+    
+    Contains both legacy trajectory format (for backward compatibility) and
+    modern v3 trace format (preferred going forward).
+    """
+    run_id: str
+    
+    # DEPRECATED: Legacy format maintained for training code compatibility.
+    # Will be removed once training migrates to reading from `trace` field.
+    # See: monorepo/trace_single_source.txt for migration plan.
+    trajectories: list[RolloutTrajectory]
+    
+    branches: dict[str, list[str]] = Field(default_factory=dict)
+    metrics: RolloutMetrics
+    aborted: bool = False
+    ops_executed: int = 0
+    
+    # OPTIONAL: correlation ID for linking rollout to inference traces
+    # If not provided, trainer will infer it from trajectory.inference_url ?cid=... parameter
+    trace_correlation_id: str | None = None
+    
+    # PREFERRED: v3 trace format (SessionTrace). This is the single source of truth
+    # for rollout data and should be used by all new code. Contains richer data than
+    # trajectories including token IDs, logprobs, timing, and multimodal content.
+    trace: dict[str, Any] | None = None
+    pipeline_metadata: dict[str, Any] = Field(default_factory=dict)
+
+
+class _ExtraAllowModel(BaseModel):
+    """Base helper that preserves unknown keys while still exposing typed attributes."""
+
+    model_config = ConfigDict(extra="allow")
+
+
+class TaskDescriptor(_ExtraAllowModel):
+    """Human-readable task identifiers shown in UIs and logs."""
+
+    id: str
+    name: str
+    description: str | None = None
+    version: str | None = None
+
+
+class DatasetInfo(_ExtraAllowModel):
+    """Metadata about the prompt/task dataset powering the environment."""
+
+    id: str | None = None
+    name: str | None = None
+    version: str | None = None
+    splits: list[str] | None = None
+    default_split: str | None = None
+    description: str | None = None
+
+
+class RubricCriterion(_ExtraAllowModel):
+    id: str
+    description: str
+    weight: float | None = None
+
+
+class RubricSection(_ExtraAllowModel):
+    name: str
+    criteria: list[RubricCriterion] = Field(default_factory=list)
+
+
+class RubricInfo(_ExtraAllowModel):
+    """Outcome and event scoring definitions used by judges."""
+
+    outcome: RubricSection | None = None
+    events: RubricSection | None = None
+
+
+class InferenceInfo(_ExtraAllowModel):
+    """Recommended defaults for policy model routing."""
+
+    model: str | None = None
+    inference_url: str | None = None
+
+
+class LimitsInfo(_ExtraAllowModel):
+    """Operational limits the environment enforces."""
+
+    max_turns: int | None = None
+    max_response_tokens: int | None = None
+    timeout_seconds: int | None = None
+
+
+class TaskInfo(_ExtraAllowModel):
+    """Static metadata describing the capabilities of a Task App task."""
+
+    task: TaskDescriptor
+    environment: str
+    dataset: DatasetInfo
+    rubric: RubricInfo
+    inference: InferenceInfo
+    limits: LimitsInfo
+    task_metadata: dict[str, Any] = Field(
+        default_factory=dict,
+        description="Task-specific extras (e.g. prompt version info, documentation links).",
+    )

synth_ai/sdk/task/datasets.py

@@ -0,0 +1,108 @@
+"""Dataset registry and helpers shared by Task Apps."""
+
+from __future__ import annotations
+
+from collections.abc import Callable, Hashable
+from typing import Any, List
+
+from pydantic import BaseModel, Field, field_validator
+
+
+class TaskDatasetSpec(BaseModel):
+    """Declarative metadata describing a dataset that a Task App exposes."""
+
+    id: str
+    name: str
+    version: str | None = None
+    splits: list[str] = Field(default_factory=list)
+    default_split: str | None = None
+    cardinality: int | None = None
+    description: str | None = None
+
+    @field_validator("default_split")
+    @classmethod
+    def _validate_default_split(cls, value: str | None, info):
+        values = info.data if hasattr(info, "data") else {}  # type: ignore[attr-defined]
+        if value and value not in (values.get("splits") or []):
+            raise ValueError("default_split must be one of splits when provided")
+        return value
+
+
+RegistryLoader = Callable[[TaskDatasetSpec], Any]
+
+
+class TaskDatasetRegistry:
+    """Lightweight registry mapping dataset specs to loader callables."""
+
+    def __init__(self) -> None:
+        self._entries: dict[str, tuple[TaskDatasetSpec, RegistryLoader, bool]] = {}
+        self._cache: dict[Hashable, Any] = {}
+
+    def register(
+        self, spec: TaskDatasetSpec, loader: RegistryLoader, *, cache: bool = True
+    ) -> None:
+        """Register a dataset loader and its metadata."""
+
+        self._entries[spec.id] = (spec, loader, cache)
+
+    def describe(self, dataset_id: str) -> TaskDatasetSpec:
+        if dataset_id not in self._entries:
+            raise KeyError(f"Dataset not registered: {dataset_id}")
+        spec, _, _ = self._entries[dataset_id]
+        return spec
+
+    def list(self) -> List[TaskDatasetSpec]:
+        return [entry[0] for entry in self._entries.values()]
+
+    def get(self, spec: TaskDatasetSpec | str) -> Any:
+        """Return dataset materialisation (with optional caching)."""
+
+        if isinstance(spec, str):
+            if spec not in self._entries:
+                raise KeyError(f"Dataset not registered: {spec}")
+            base_spec, loader, cache_enabled = self._entries[spec]
+            effective_spec = base_spec
+        else:
+            if spec.id not in self._entries:
+                raise KeyError(f"Dataset not registered: {spec.id}")
+            base_spec, loader, cache_enabled = self._entries[spec.id]
+            effective_spec = base_spec.model_copy(update=spec.model_dump(exclude_unset=True))
+
+        cache_key: Hashable = (
+            effective_spec.id,
+            effective_spec.version,
+            effective_spec.default_split,
+        )
+        if cache_enabled:
+            if cache_key not in self._cache:
+                self._cache[cache_key] = loader(effective_spec)
+            return self._cache[cache_key]
+        return loader(effective_spec)
+
+    @staticmethod
+    def ensure_split(spec: TaskDatasetSpec, split: str | None) -> str:
+        """Validate that `split` exists on the spec; return a concrete split."""
+
+        if not spec.splits:
+            return split or spec.default_split or "default"
+        if split is None:
+            if spec.default_split:
+                return spec.default_split
+            raise ValueError(f"split must be provided for dataset {spec.id}")
+        if split not in spec.splits:
+            raise ValueError(f"Unknown split '{split}' for dataset {spec.id}")
+        return split
+
+    @staticmethod
+    def normalise_seed(seed: Any, *, cardinality: int | None = None) -> int:
+        """Normalise arbitrary seed input into a bounded non-negative integer."""
+
+        try:
+            value = int(seed)
+        except Exception as exc:  # pragma: no cover - defensive
+            raise ValueError(f"Seed must be convertible to int (got {seed!r})") from exc
+        if value < 0:
+            value = abs(value)
+        if cardinality and cardinality > 0:
+            value = value % cardinality
+        return value