synth-ai 0.2.9.dev0__py3-none-any.whl → 0.2.23.dev3__py3-none-any.whl

synth_ai/cli/commands/train/judge_schemas.py

@@ -0,0 +1,200 @@
+"""
+Pydantic schemas for judge/rubric configuration.
+
+These models define the ACTUAL fields used by the backend judge service,
+with all dead code removed. This is the single source of truth for what
+gets sent in HTTP requests.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Optional
+
+from pydantic import Field, model_validator
+from synth_ai.api.train.configs.shared import ExtraModel
+
+__all__ = [
+    "RubricWeightsConfig",
+    "RubricConfig",
+    "JudgeOptionsConfig",
+    "JudgeConfig",
+    "JudgeRequestPayload",
+]
+
+
+class RubricWeightsConfig(ExtraModel):
+    """
+    Reward blending weights (client-side only, not sent to backend).
+    
+    These weights control how env rewards, event judge scores, and outcome
+    judge scores are combined into a final reward signal for policy gradients.
+    
+    Formula:
+        total_reward = (env * env_return) + (event * sum(event_scores)) + (outcome * outcome_score)
+    """
+    env: float = Field(
+        default=1.0,
+        description="Weight for environment rewards (task app native rewards)",
+        ge=0.0,
+    )
+    event: float = Field(
+        default=0.0,
+        description="Weight for per-event judge scores (step-level judging)",
+        ge=0.0,
+    )
+    outcome: float = Field(
+        default=0.0,
+        description="Weight for outcome judge score (episode-level judging)",
+        ge=0.0,
+    )
+
+    @model_validator(mode="after")
+    def _validate_weights_sum(self) -> RubricWeightsConfig:
+        """Ensure at least one weight is non-zero."""
+        if self.env == 0.0 and self.event == 0.0 and self.outcome == 0.0:
+            raise ValueError("At least one reward weight must be non-zero")
+        return self
+
+
+class RubricConfig(ExtraModel):
+    """
+    Top-level rubric configuration.
+    
+    Controls whether rubric-based judging is enabled and how rewards are blended.
+    """
+    enabled: bool = Field(
+        default=False,
+        description="Master switch for rubric-based judging",
+    )
+    weights: RubricWeightsConfig = Field(
+        default_factory=RubricWeightsConfig,
+        description="Reward blending weights (env/event/outcome)",
+    )
+
+
+class JudgeOptionsConfig(ExtraModel):
+    """
+    Judge provider options (sent to backend in HTTP request).
+    
+    These fields are sent in the "options" object of the judge score request.
+    All fields here map directly to the backend JudgeOptions schema.
+    """
+    provider: str = Field(
+        ...,
+        description="Judge provider type ('openai', 'groq', 'gemini')",
+        pattern=r"^(openai|groq|gemini)$",
+    )
+    model: str = Field(
+        ...,
+        description="Model identifier (e.g., 'openai/gpt-oss-120b', 'gpt-5')",
+        min_length=1,
+    )
+    rubric_id: Optional[str] = Field(
+        default=None,
+        description="Base rubric identifier (e.g., 'crafter/bundle@v1')",
+    )
+    event: bool = Field(
+        default=True,
+        description="Enable per-event (step-level) judging",
+    )
+    outcome: bool = Field(
+        default=True,
+        description="Enable outcome (episode-level) judging",
+    )
+    timeout_s: Optional[float] = Field(
+        default=None,
+        description="Request timeout in seconds",
+        gt=0,
+    )
+    metadata: dict[str, Any] = Field(
+        default_factory=dict,
+        description="Optional metadata (e.g., {'async': true, 'custom_field': 'value'})",
+    )
+    rubric_overrides: dict[str, Any] = Field(
+        default_factory=dict,
+        description=(
+            "Static rubric criteria overrides (rarely used - TaskInfo overrides take priority). "
+            "Format: {'event': {'criteria': [...]}, 'outcome': {'criteria': [...]}}"
+        ),
+    )
+
+    @model_validator(mode="after")
+    def _validate_at_least_one_enabled(self) -> JudgeOptionsConfig:
+        """Ensure at least one judging type is enabled."""
+        if not self.event and not self.outcome:
+            raise ValueError("At least one of 'event' or 'outcome' must be enabled")
+        return self
+
+
+class JudgeConfig(ExtraModel):
+    """
+    Top-level judge configuration.
+    
+    This is parsed from TOML [judge] section and contains all judge-related settings.
+    """
+    options: JudgeOptionsConfig = Field(
+        ...,
+        description="Judge provider options (sent to backend)",
+    )
+
+
+# HTTP Request Payload Structures (for documentation/type safety)
+
+class JudgeRequestPayload(ExtraModel):
+    """
+    HTTP request payload structure for POST /api/judge/v1/score.
+    
+    This is the ACTUAL payload sent to the backend judge service.
+    Used for type safety and documentation only.
+    """
+    policy_name: str = Field(..., description="Name of the policy being evaluated")
+    task_app: dict[str, Any] = Field(..., description="Task app metadata (id, base_url)")
+    trace: dict[str, Any] = Field(..., description="Tracing v3 payload (event_history, metadata)")
+    options: dict[str, Any] = Field(..., description="Judge options (provider, model, etc.)")
+
+    class Config:
+        extra = "allow"  # Backend might add extra fields
+
+
+# Helper to convert to backend request format
+
+def build_judge_http_options(
+    options_config: JudgeOptionsConfig,
+    *,
+    rubric_overrides_from_task_info: Optional[dict[str, Any]] = None,
+) -> dict[str, Any]:
+    """
+    Build the 'options' dict for HTTP request to backend judge.
+    
+    Args:
+        options_config: Validated judge options from TOML
+        rubric_overrides_from_task_info: Dynamic overrides fetched from TaskInfo (takes priority)
+    
+    Returns:
+        Dict ready to send in HTTP request payload
+    """
+    payload = {
+        "provider": options_config.provider,
+        "model": options_config.model,
+        "event": options_config.event,
+        "outcome": options_config.outcome,
+    }
+    
+    # Optional fields
+    if options_config.rubric_id:
+        payload["rubric_id"] = options_config.rubric_id
+    
+    if options_config.timeout_s is not None:
+        payload["timeout_s"] = options_config.timeout_s
+    
+    if options_config.metadata:
+        payload["metadata"] = options_config.metadata
+    
+    # Rubric overrides: TaskInfo takes priority over static config
+    if rubric_overrides_from_task_info:
+        payload["rubric_overrides"] = rubric_overrides_from_task_info
+    elif options_config.rubric_overrides:
+        payload["rubric_overrides"] = options_config.rubric_overrides
+    
+    return payload
+

synth_ai/cli/commands/train/judge_validation.py

@@ -0,0 +1,305 @@
+"""
+Validation logic for judge/rubric configuration from TOML.
+
+This module validates and normalizes judge/rubric config, removing all dead fields
+and ensuring only the fields actually used by the backend are present.
+"""
+
+from __future__ import annotations
+
+import warnings
+from collections.abc import MutableMapping
+from typing import Any, Optional, Tuple
+
+from pydantic import ValidationError
+
+from .errors import InvalidJudgeConfigError, InvalidRubricConfigError
+from .judge_schemas import JudgeConfig, JudgeOptionsConfig, RubricConfig, RubricWeightsConfig
+
+__all__ = [
+    "validate_judge_config",
+    "validate_rubric_config",
+    "extract_and_validate_judge_rubric",
+]
+
+# Dead fields that should trigger deprecation warnings
+DEPRECATED_RUBRIC_FIELDS = {
+    "model",
+    "api_base",
+    "api_key_env",
+    "event",
+    "outcome",
+}
+
+DEPRECATED_JUDGE_FIELDS = {
+    "type",
+    "timeout_s",  # Moved to judge.options.timeout_s
+}
+
+DEPRECATED_JUDGE_OPTIONS_FIELDS = {
+    "max_concurrency",
+    "tracks",
+}
+
+
+def _warn_deprecated_fields(section: str, fields: set[str], present_fields: set[str]) -> None:
+    """Emit deprecation warnings for dead fields that are present in config."""
+    deprecated_present = fields & present_fields
+    if deprecated_present:
+        field_list = ", ".join(sorted(deprecated_present))
+        warnings.warn(
+            f"[{section}] contains deprecated fields that are no longer used: {field_list}. "
+            f"These fields will be ignored and should be removed from your config. "
+            f"See judge/rubric cleanup guide for details.",
+            DeprecationWarning,
+            stacklevel=3,
+        )
+
+
+def validate_rubric_config(config: MutableMapping[str, Any]) -> RubricConfig:
+    """
+    Validate and normalize rubric configuration from TOML.
+    
+    Args:
+        config: Raw [rubric] section from TOML
+        
+    Returns:
+        Validated RubricConfig instance
+        
+    Raises:
+        InvalidRubricConfigError: If validation fails
+    """
+    if not config:
+        # Default: rubric disabled
+        return RubricConfig(enabled=False)
+    
+    config_dict = dict(config)
+    
+    # Warn about deprecated fields
+    _warn_deprecated_fields("rubric", DEPRECATED_RUBRIC_FIELDS, set(config_dict.keys()))
+    
+    # Warn about deprecated subsections
+    if "event" in config_dict:
+        warnings.warn(
+            "[rubric.event] section is deprecated and no longer used. "
+            "Criteria are now fetched dynamically from TaskInfo or specified in "
+            "[judge.options.rubric_overrides]. This section will be ignored.",
+            DeprecationWarning,
+            stacklevel=2,
+        )
+    
+    if "outcome" in config_dict:
+        warnings.warn(
+            "[rubric.outcome] section is deprecated and no longer used. "
+            "Criteria are now fetched dynamically from TaskInfo or specified in "
+            "[judge.options.rubric_overrides]. This section will be ignored.",
+            DeprecationWarning,
+            stacklevel=2,
+        )
+    
+    # Extract only valid fields
+    enabled = config_dict.get("enabled", False)
+    weights_dict = config_dict.get("weights", {})
+    
+    # Validate using Pydantic
+    try:
+        if not isinstance(weights_dict, dict):
+            raise ValueError("[rubric.weights] must be a dictionary")
+        
+        weights = RubricWeightsConfig(**weights_dict)
+        return RubricConfig(enabled=enabled, weights=weights)
+    
+    except ValidationError as exc:
+        errors = []
+        for error in exc.errors():
+            loc = ".".join(str(x) for x in error["loc"])
+            msg = error["msg"]
+            errors.append(f"  • rubric.{loc}: {msg}")
+        raise InvalidRubricConfigError(
+            detail="Rubric validation failed:\n" + "\n".join(errors)
+        ) from exc
+    except Exception as exc:
+        raise InvalidRubricConfigError(
+            detail=f"Rubric validation failed: {exc}"
+        ) from exc
+
+
+def validate_judge_config(config: MutableMapping[str, Any]) -> Optional[JudgeConfig]:
+    """
+    Validate and normalize judge configuration from TOML.
+    
+    Args:
+        config: Raw [judge] section from TOML
+        
+    Returns:
+        Validated JudgeConfig instance, or None if not present
+        
+    Raises:
+        InvalidJudgeConfigError: If validation fails
+    """
+    if not config:
+        return None
+    
+    config_dict = dict(config)
+    
+    # Warn about deprecated top-level fields
+    _warn_deprecated_fields("judge", DEPRECATED_JUDGE_FIELDS, set(config_dict.keys()))
+    
+    # Extract judge.options (required)
+    options_dict = config_dict.get("options")
+    if not options_dict:
+        raise InvalidJudgeConfigError(
+            detail="[judge.options] section is required when [judge] is present"
+        )
+    
+    if not isinstance(options_dict, dict):
+        raise InvalidJudgeConfigError(
+            detail="[judge.options] must be a dictionary"
+        )
+    
+    # Warn about deprecated options fields
+    _warn_deprecated_fields(
+        "judge.options",
+        DEPRECATED_JUDGE_OPTIONS_FIELDS,
+        set(options_dict.keys()),
+    )
+    
+    # Remove deprecated fields from options
+    options_dict = {
+        k: v for k, v in options_dict.items()
+        if k not in DEPRECATED_JUDGE_OPTIONS_FIELDS
+    }
+    
+    # Migrate judge.timeout_s to judge.options.timeout_s if present
+    if "timeout_s" in config_dict and "timeout_s" not in options_dict:
+        warnings.warn(
+            "[judge].timeout_s is deprecated. Use [judge.options].timeout_s instead. "
+            "Auto-migrating for now.",
+            DeprecationWarning,
+            stacklevel=2,
+        )
+        options_dict["timeout_s"] = config_dict["timeout_s"]
+    
+    # Validate using Pydantic
+    try:
+        options = JudgeOptionsConfig(**options_dict)
+        return JudgeConfig(options=options)
+    
+    except ValidationError as exc:
+        errors = []
+        for error in exc.errors():
+            loc = ".".join(str(x) for x in error["loc"])
+            msg = error["msg"]
+            errors.append(f"  • judge.options.{loc}: {msg}")
+        raise InvalidJudgeConfigError(
+            detail="Judge validation failed:\n" + "\n".join(errors)
+        ) from exc
+    except Exception as exc:
+        raise InvalidJudgeConfigError(
+            detail=f"Judge validation failed: {exc}"
+        ) from exc
+
+
+def extract_and_validate_judge_rubric(
+    toml_config: MutableMapping[str, Any]
+) -> Tuple[RubricConfig, Optional[JudgeConfig]]:
+    """
+    Extract and validate judge/rubric config from full TOML config.
+    
+    Args:
+        toml_config: Full TOML configuration dict
+        
+    Returns:
+        Tuple of (validated_rubric, validated_judge_or_none)
+        
+    Raises:
+        InvalidRubricConfigError: If rubric validation fails
+        InvalidJudgeConfigError: If judge validation fails
+    """
+    rubric_dict = toml_config.get("rubric", {})
+    judge_dict = toml_config.get("judge", {})
+    
+    # Validate rubric
+    rubric_config = validate_rubric_config(rubric_dict)
+    
+    # Validate judge (if present)
+    judge_config = validate_judge_config(judge_dict) if judge_dict else None
+    
+    # Cross-validation: If rubric is enabled, judge options should be present
+    if rubric_config.enabled and not judge_config:
+        warnings.warn(
+            "[rubric].enabled=true but [judge] section is missing. "
+            "Rubric-based judging requires judge configuration. "
+            "Rubric scoring will be disabled.",
+            UserWarning,
+            stacklevel=2,
+        )
+        rubric_config.enabled = False
+    
+    # Cross-validation: Warn if weights don't align with enabled judging types
+    if rubric_config.enabled and judge_config:
+        weights = rubric_config.weights
+        options = judge_config.options
+        
+        if weights.event > 0 and not options.event:
+            warnings.warn(
+                "[rubric.weights].event > 0 but [judge.options].event=false. "
+                "Event-level judge scores will be 0 (no event judging enabled).",
+                UserWarning,
+                stacklevel=2,
+            )
+        
+        if weights.outcome > 0 and not options.outcome:
+            warnings.warn(
+                "[rubric.weights].outcome > 0 but [judge.options].outcome=false. "
+                "Outcome judge score will be 0 (no outcome judging enabled).",
+                UserWarning,
+                stacklevel=2,
+            )
+    
+    return rubric_config, judge_config
+
+
+# Helper to check if config has any deprecated fields (for testing/migration)
+
+def check_for_deprecated_fields(toml_config: MutableMapping[str, Any]) -> dict[str, list[str]]:
+    """
+    Check TOML config for deprecated fields without validation.
+    
+    Returns dict of {section: [deprecated_field_names]} for reporting.
+    """
+    deprecated: dict[str, list[str]] = {}
+    
+    rubric_dict = toml_config.get("rubric", {})
+    if rubric_dict:
+        found = [
+            field for field in DEPRECATED_RUBRIC_FIELDS
+            if field in rubric_dict
+        ]
+        if "event" in rubric_dict:
+            found.append("event (entire section)")
+        if "outcome" in rubric_dict:
+            found.append("outcome (entire section)")
+        if found:
+            deprecated["rubric"] = found
+    
+    judge_dict = toml_config.get("judge", {})
+    if judge_dict:
+        found = [
+            field for field in DEPRECATED_JUDGE_FIELDS
+            if field in judge_dict
+        ]
+        if found:
+            deprecated["judge"] = found
+        
+        options_dict = judge_dict.get("options", {})
+        if options_dict:
+            options_found = [
+                field for field in DEPRECATED_JUDGE_OPTIONS_FIELDS
+                if field in options_dict
+            ]
+            if options_found:
+                deprecated["judge.options"] = options_found
+    
+    return deprecated
+