Datasets:
Download src/explicit_learning/config.py from sungguk/visual-answerability: direct link, hf CLI and curl.
- Browser
- Download file 14.7 kB
-
https://huggingface.co/datasets/sungguk/visual-answerability/resolve/main/src/explicit_learning/config.py
- Command line
-
hf download hf://datasets/sungguk/visual-answerability/src/explicit_learning/config.py
-
curl -L -o config.py https://huggingface.co/datasets/sungguk/visual-answerability/resolve/main/src/explicit_learning/config.py
14.7 kB
| """Pydantic models for the resource manifest and experiment defaults. | |
| These models are the runtime source of truth for ``configs/resources.yaml`` and | |
| ``configs/experiment.yaml``. They enforce the research invariants that | |
| ``scripts/validate_spec.py`` checks statically, so a config that loads here is | |
| safe to use in a run. Floating revisions, MathVista in training, invented MMK12 | |
| licenses, and a malformed run matrix all fail at load time. | |
| """ | |
| from __future__ import annotations | |
| import math | |
| from pathlib import Path | |
| from typing import Any | |
| import yaml | |
| from pydantic import BaseModel, ConfigDict, Field, ValidationError, model_validator | |
| from .hashing import canonical_config_hash, is_git_revision | |
| # Revision policy: every HF/Git resource must resolve to an exact 40-hex commit. | |
| # ADR-0002: gold authority is a machine-checkable executor certificate; no | |
| # per-example human edit/label/review/adjudication is permitted. | |
| EXPECTED_MAIN_RUNS = 9 | |
| EXPECTED_RUN_COUNTS: dict[tuple[str, str], int] = { | |
| ("qwen35_2b", "answer_grpo"): 1, | |
| ("qwen35_2b", "papo_controlled"): 2, | |
| ("qwen35_2b", "defacto_controlled"): 2, | |
| ("qwen35_2b", "intervention_grpo"): 2, | |
| ("qwen35_2b", "evi_po"): 2, | |
| } | |
| # Core ingredient isolation: remove each EVI-PO auxiliary loss once. Broader | |
| # target/PAPO/margin sweeps remain opt-in rather than an automatic GPU queue. | |
| EXPECTED_ABLATION_RUNS = 2 | |
| EXPECTED_ABLATION_ARMS: tuple[str, ...] = ( | |
| "evi_po_no_direction", | |
| "evi_po_no_evidence", | |
| ) | |
| class ConfigError(ValueError): | |
| """Raised when a config file violates a research invariant.""" | |
| def _load_yaml(path: str | Path) -> dict[str, Any]: | |
| text = Path(path).read_text(encoding="utf-8") | |
| data = yaml.safe_load(text) | |
| if not isinstance(data, dict): | |
| raise ConfigError(f"{path} must contain a top-level mapping") | |
| return data | |
| class ResourceModel(BaseModel): | |
| model_config = ConfigDict(extra="allow") | |
| role: str | |
| provider: str | |
| repo_id: str | |
| revision: str | |
| url: str | |
| license: str | None = None | |
| access: str = "public" | |
| estimated_size_bytes: int | None = None | |
| use: str | None = None | |
| def _check_revision(self) -> ResourceModel: | |
| if not is_git_revision(self.revision): | |
| raise ConfigError( | |
| f"models.{self.repo_id}.revision must be a 40-hex commit, got {self.revision!r}" | |
| ) | |
| return self | |
| class ResourceDataset(BaseModel): | |
| model_config = ConfigDict(extra="allow") | |
| role: str | |
| provider: str | |
| repo_id: str | |
| revision: str | |
| url: str | |
| declared_license: str | None = None | |
| policy: str = "" | |
| configs: list[str] | None = None | |
| # Split metadata is heterogeneous across sources (flat counts for most, | |
| # per-config counts for MMMU), so values are left loosely typed; specific | |
| # guards in validate_against_resources check exact values where it matters. | |
| splits: dict[str, Any] | None = None | |
| splits_per_config: dict[str, Any] | None = None | |
| notes: list[str] | None = None | |
| def _check_revision(self) -> ResourceDataset: | |
| if not is_git_revision(self.revision): | |
| raise ConfigError( | |
| f"datasets.{self.repo_id}.revision must be a 40-hex commit, got {self.revision!r}" | |
| ) | |
| return self | |
| class ResourceCheckpoint(BaseModel): | |
| model_config = ConfigDict(extra="allow") | |
| role: str | |
| provider: str | |
| repo_id: str | |
| revision: str | |
| url: str | |
| policy: str = "" | |
| def _check_revision(self) -> ResourceCheckpoint: | |
| if not is_git_revision(self.revision): | |
| raise ConfigError( | |
| f"checkpoints.{self.repo_id}.revision must be a 40-hex commit, " | |
| f"got {self.revision!r}" | |
| ) | |
| return self | |
| class ResourceRepository(BaseModel): | |
| model_config = ConfigDict(extra="allow") | |
| url: str | |
| revision: str | |
| role: str | |
| def _check_revision(self) -> ResourceRepository: | |
| if not is_git_revision(self.revision): | |
| raise ConfigError( | |
| f"repositories.{self.url}.revision must be a 40-hex commit, got {self.revision!r}" | |
| ) | |
| return self | |
| class ResourcesManifest(BaseModel): | |
| model_config = ConfigDict(extra="allow") | |
| manifest_version: int | |
| models: dict[str, ResourceModel] = Field(default_factory=dict) | |
| datasets: dict[str, ResourceDataset] = Field(default_factory=dict) | |
| checkpoints: dict[str, ResourceCheckpoint] = Field(default_factory=dict) | |
| repositories: dict[str, ResourceRepository] = Field(default_factory=dict) | |
| # C1 source-native structured worlds (PlotQA, Geometry3K) live outside the HF | |
| # ``datasets`` map: they are fetched from pinned GitHub revisions and feed the | |
| # certificate pipeline directly (docs/02 §4-6). Kept loosely typed here; the | |
| # adapter layer owns their detailed schema. | |
| structured_sources: dict[str, Any] = Field(default_factory=dict) | |
| def model(self, name: str) -> ResourceModel: | |
| if name not in self.models: | |
| raise ConfigError(f"unknown model resource {name!r}") | |
| return self.models[name] | |
| def dataset(self, name: str) -> ResourceDataset: | |
| if name not in self.datasets: | |
| raise ConfigError(f"unknown dataset resource {name!r}") | |
| return self.datasets[name] | |
| class ExperimentRun(BaseModel): | |
| model_config = ConfigDict(extra="forbid") | |
| run_id: str | |
| model: str | |
| arm: str | |
| seed: int | |
| class CommonStudent(BaseModel): | |
| model_config = ConfigDict(extra="allow") | |
| # One trainable 2B student. The v1 C1 release uses no teacher/verifier model. | |
| model_resource: str | |
| precision: str = "bf16" | |
| attention_implementation: str = "flash_attention_2" | |
| class ExperimentData(BaseModel): | |
| model_config = ConfigDict(extra="allow") | |
| train_sources: dict[str, Any] = Field(default_factory=dict) | |
| evaluation: dict[str, Any] = Field(default_factory=dict) | |
| class ExperimentConfig(BaseModel): | |
| model_config = ConfigDict(extra="allow") | |
| experiment_schema_version: int | |
| study_name: str | |
| reproducibility: dict[str, Any] = Field(default_factory=dict) | |
| labels: dict[str, Any] = Field(default_factory=dict) | |
| data: ExperimentData = Field(default_factory=ExperimentData) | |
| # Automated release/audit contract. V1 requires zero per-example model calls; | |
| # Terra/Luna are one-time diagnostics and never admission/gold authorities. | |
| automated_compilation_and_audit: dict[str, Any] | |
| common_student: CommonStudent | |
| rewards: dict[str, Any] = Field(default_factory=dict) | |
| main_rl_runs: list[ExperimentRun] | |
| # Ingredient-isolation suite (see EXPECTED_ABLATION_ARMS). Numeric knobs | |
| # (lambda_direction, mask_ratio, ...) live in code (rewards.ABLATION_ARM_KNOBS); | |
| # the manifest only pins identities. | |
| ablation_rl_runs: list[ExperimentRun] = Field(default_factory=list) | |
| sft_ablation_runs: list[dict[str, Any]] = Field(default_factory=list) | |
| def _validate_run_matrix(self) -> ExperimentConfig: | |
| runs = self.main_rl_runs | |
| if len(runs) != EXPECTED_MAIN_RUNS: | |
| raise ConfigError( | |
| f"main_rl_runs must contain {EXPECTED_MAIN_RUNS} runs, found {len(runs)}" | |
| ) | |
| run_ids = [run.run_id for run in runs] | |
| if len(run_ids) != len(set(run_ids)): | |
| duplicates = sorted({rid for rid in run_ids if run_ids.count(rid) > 1}) | |
| raise ConfigError(f"duplicate run_id in main_rl_runs: {duplicates}") | |
| counts: dict[tuple[str, str], int] = {} | |
| for run in runs: | |
| key = (run.model, run.arm) | |
| counts[key] = counts.get(key, 0) + 1 | |
| if counts != EXPECTED_RUN_COUNTS: | |
| raise ConfigError(f"unexpected main RL run matrix: {counts}") | |
| evi_po = self.rewards.get("evi_po") | |
| if not isinstance(evi_po, dict): | |
| raise ConfigError("rewards.evi_po must be an object") | |
| if evi_po.get("required_relationships") != ["FULL", "CONTROL", "MISSING"]: | |
| raise ConfigError("EVI-PO requires FULL, CONTROL, and MISSING relationships") | |
| if evi_po.get("optional_relationships") != ["SUBSTITUTE"]: | |
| raise ConfigError("EVI-PO SUBSTITUTE relationship must remain optional") | |
| for name in ("lambda_direction", "lambda_evidence", "margin"): | |
| value = evi_po.get(name) | |
| if ( | |
| isinstance(value, bool) | |
| or not isinstance(value, int | float) | |
| or not math.isfinite(float(value)) | |
| or value < 0 | |
| ): | |
| raise ConfigError(f"rewards.evi_po.{name} must be non-negative") | |
| if evi_po.get("require_usable_response_to_image_attention") is not True: | |
| raise ConfigError("EVI-PO must require usable response-to-image attention") | |
| self._validate_ablation_matrix() | |
| return self | |
| def _validate_ablation_matrix(self) -> None: | |
| """Enforce the two direct EVI objective-removal runs.""" | |
| runs = self.ablation_rl_runs | |
| if len(runs) != EXPECTED_ABLATION_RUNS: | |
| raise ConfigError( | |
| f"ablation_rl_runs must contain {EXPECTED_ABLATION_RUNS} runs, " | |
| f"found {len(runs)}" | |
| ) | |
| run_ids = [run.run_id for run in runs] | |
| if len(run_ids) != len(set(run_ids)): | |
| duplicates = sorted({rid for rid in run_ids if run_ids.count(rid) > 1}) | |
| raise ConfigError(f"duplicate run_id in ablation_rl_runs: {duplicates}") | |
| identities = {(run.model, run.arm, run.seed) for run in runs} | |
| expected = {(model, arm, seed) for arm in EXPECTED_ABLATION_ARMS | |
| for model in ("qwen35_2b",) for seed in (1,)} | |
| if identities != expected: | |
| raise ConfigError(f"unexpected ablation RL run matrix: {sorted(identities)}") | |
| def validate_against_resources(self, resources: ResourcesManifest) -> None: | |
| """Enforce cross-file invariants between experiment and resources.""" | |
| name = self.common_student.model_resource | |
| if name not in resources.models: | |
| raise ConfigError(f"common_student.model_resource references unknown model {name!r}") | |
| aca = self.automated_compilation_and_audit | |
| if aca.get("total_reasoning_vlm_calls_for_data_generation") != 0: | |
| raise ConfigError("v1 data generation must use exactly zero reasoning VLM calls") | |
| if aca.get("total_ocr_vlm_calls_for_data_generation") != 0: | |
| raise ConfigError("v1 data generation must use exactly zero OCR VLM calls") | |
| if aca.get("gold_authority") != ( | |
| "deterministic_primary_and_reference_executor_certificates" | |
| ): | |
| raise ConfigError("v1 gold authority must be deterministic executor certificates") | |
| # Train sources may be HF datasets (C2) or GitHub structured worlds (C1, | |
| # e.g. PlotQA/Geometry3K); either is acceptable. | |
| for name in self.data.train_sources: | |
| if name not in resources.datasets and name not in resources.structured_sources: | |
| raise ConfigError(f"train source {name!r} is absent from resources") | |
| if "mathvista" in self.data.train_sources: | |
| raise ConfigError("MathVista must never appear in train_sources") | |
| for name in self.data.evaluation.get("certified_intervention", {}): | |
| if name not in resources.datasets and name not in resources.structured_sources: | |
| raise ConfigError( | |
| f"certified_intervention source {name!r} is absent from resources" | |
| ) | |
| for name in self.data.evaluation.get("untouched", []): | |
| if name not in resources.datasets: | |
| raise ConfigError(f"untouched source {name!r} is absent from resources.yaml") | |
| # Dataset-specific guards mirrored from validate_spec.py. | |
| mv = resources.datasets.get("mathvista") | |
| if mv and mv.policy != "strict_evaluation_only_training_prohibited_by_card": | |
| raise ConfigError("MathVista policy must remain strict evaluation-only") | |
| mmk = resources.datasets.get("mmk12") | |
| if mmk and mmk.declared_license is not None: | |
| raise ConfigError("MMK12 must not be assigned an invented license") | |
| chart = resources.datasets.get("chartqa") | |
| if chart: | |
| splits = chart.splits or {} | |
| if splits.get("val") != 1920 or "validation" in splits: | |
| raise ConfigError("ChartQA must use its actual split name 'val'") | |
| mmmu_pro = resources.datasets.get("mmmu_pro") | |
| if mmmu_pro and mmmu_pro.configs != [ | |
| "standard (10 options)", | |
| "standard (4 options)", | |
| "vision", | |
| ]: | |
| raise ConfigError("MMMU-Pro config names must match Hugging Face metadata exactly") | |
| def canonical_hash(self) -> str: | |
| """Machine/path-independent hash of this experiment config.""" | |
| return canonical_config_hash(self.model_dump(mode="python")) | |
| def _format_validation_error(exc: ValidationError) -> str: | |
| """Render a pydantic ``ValidationError`` as a compact, human-readable string. | |
| Validators raise :class:`ConfigError` (a ``ValueError``), which pydantic | |
| wraps into a ``ValidationError``; we surface the original messages here so | |
| callers see the research-invariant reason rather than pydantic's envelope. | |
| """ | |
| parts: list[str] = [] | |
| for err in exc.errors(): | |
| loc = ".".join(str(part) for part in err.get("loc", ())) | |
| msg = err.get("msg", "") | |
| if msg.startswith("Value error, "): | |
| msg = msg[len("Value error, ") :] | |
| parts.append(f"{loc}: {msg}" if loc else msg) | |
| return "; ".join(parts) or str(exc) | |
| def load_resources(path: str | Path) -> ResourcesManifest: | |
| try: | |
| return ResourcesManifest.model_validate(_load_yaml(path)) | |
| except ValidationError as exc: | |
| raise ConfigError(_format_validation_error(exc)) from exc | |
| def load_experiment(path: str | Path) -> ExperimentConfig: | |
| try: | |
| return ExperimentConfig.model_validate(_load_yaml(path)) | |
| except ValidationError as exc: | |
| raise ConfigError(_format_validation_error(exc)) from exc | |
| def load_config_pair( | |
| resources_path: str | Path, | |
| experiment_path: str | Path, | |
| ) -> tuple[ResourcesManifest, ExperimentConfig]: | |
| """Load both configs and enforce their cross-file invariants.""" | |
| resources = load_resources(resources_path) | |
| experiment = load_experiment(experiment_path) | |
| experiment.validate_against_resources(resources) | |
| return resources, experiment | |