sungguk's picture
Release visual answerability benchmark v1.0.0
e1ced61 verified
Raw
History Blame Contribute Delete
21.8 kB
"""Frozen training config and immutable planned-run manifest builders."""
from __future__ import annotations
import json
import math
from dataclasses import asdict, dataclass
from datetime import datetime
from pathlib import Path
from typing import Any, Literal
from ..atomic_io import atomic_write_json
from ..hashing import canonical_json, canonical_json_hash, is_git_revision, is_sha256
from .evi_po_contract import EVI_PO_ADAPTER_VERSION
from .papo_contract import (
PAPO_ADAPTER_VERSION,
PAPO_UPSTREAM_SOURCE_SHA256,
PAPO_UPSTREAM_SOURCE_URL,
)
from .rewards import arm_trainer_kind, canonical_arm
TrainerKind = Literal["sft", "grpo", "papo", "evi_po"]
RunMode = Literal["sft", "smoke", "main"]
QWEN35_2B_REVISION = "15852e8c16360a2fea060d615a32b45270f8a8fc"
TRAINING_CONTRACT_VERSION = "evi_v2_constrained_answer_256_v1"
# Constrained decoding removes irrelevant formatting/runaway-generation noise
# while preserving every certified answer, including the public abstention
# token. vLLM completes the regex and emits EOS; the 256-token cap remains a
# hard safety boundary rather than a routinely reached training event.
ANSWER_SCHEMA_REGEX = r"<answer>(<UNANSWERABLE>|[^<>\r\n]{1,128})</answer>"
class RunArtifactError(RuntimeError):
"""Raised when a frozen run identity or write-once artifact drifts."""
@dataclass(frozen=True)
class RuntimeTrainingConfig:
"""Fully resolved values required before a GPU launcher may run."""
run_id: str
trainer_kind: TrainerKind
arm: str
seed: int
model_path: str
model_revision: str
model_snapshot_sha256: str
dataset_path: str
dataset_asset_root: str
output_dir: str
initial_checkpoint_path: str
environment_lock_path: str
environment_lock_sha256: str
evaluation_manifest_path: str
evaluation_manifest_sha256: str
system_prompt_sha256: str
precision: str
attention_implementation: str
max_prompt_tokens: int
max_completion_tokens: int
total_context_tokens: int
per_device_train_batch_size: int
gradient_accumulation_steps: int
world_size: int
generations_per_prompt: int
checkpoint_interval: int
max_optimizer_steps: int
max_completion_tokens_per_run: int
learning_rate: float
lora_rank: int
lora_alpha: int
lora_dropout: float
lora_target_modules: tuple[str, ...]
comparison_slot_manifest_sha256: str | None = None
backend_entrypoint: str | None = None
use_vllm: bool = True
run_mode: RunMode = "main"
compatibility_gate_path: str | None = None
evi_direction_loss_weight: float = 0.0
evi_evidence_loss_weight: float = 0.0
evi_direction_margin: float = 0.0
papo_mask_ratio: float = 0.6
papo_perception_loss_weight: float = 0.02
def __post_init__(self) -> None:
if not isinstance(self.run_id, str) or not self.run_id:
raise RunArtifactError("run_id must be non-empty")
if self.trainer_kind not in {"sft", "grpo", "papo", "evi_po"}:
raise RunArtifactError(f"unsupported trainer kind: {self.trainer_kind}")
if self.run_mode not in {"sft", "smoke", "main"}:
raise RunArtifactError(f"unsupported run mode: {self.run_mode}")
if self.trainer_kind == "sft" and self.run_mode != "sft":
raise RunArtifactError("SFT trainer requires run_mode=sft")
if self.trainer_kind != "sft" and self.run_mode not in {"smoke", "main"}:
raise RunArtifactError("RL trainer requires run_mode=smoke or main")
if self.trainer_kind != "sft":
try:
canonical = canonical_arm(self.arm)
except ValueError as exc:
raise RunArtifactError(f"unsupported RL arm: {self.arm}") from exc
if canonical != self.arm:
object.__setattr__(self, "arm", canonical)
if self.trainer_kind != "sft" and self.trainer_kind != arm_trainer_kind(self.arm):
raise RunArtifactError(
f"trainer kind {self.trainer_kind!r} does not match arm {self.arm!r}"
)
evi_values = (
self.evi_direction_loss_weight,
self.evi_evidence_loss_weight,
self.evi_direction_margin,
)
if any(
isinstance(value, bool)
or not isinstance(value, int | float)
or not math.isfinite(float(value))
or value < 0
for value in evi_values
):
raise RunArtifactError("EVI-PO weights and margin must be finite and non-negative")
if self.trainer_kind != "evi_po" and any(value != 0.0 for value in evi_values):
raise RunArtifactError("non-EVI trainers may not carry EVI-PO objective values")
if self.trainer_kind != "sft":
if not self.comparison_slot_manifest_sha256:
raise RunArtifactError(
"RL runs require comparison_slot_manifest_sha256 for schedule fairness"
)
if not is_sha256(self.comparison_slot_manifest_sha256):
raise RunArtifactError("comparison_slot_manifest_sha256 must be a SHA-256")
elif self.comparison_slot_manifest_sha256 is not None and not is_sha256(
self.comparison_slot_manifest_sha256
):
raise RunArtifactError("comparison_slot_manifest_sha256 must be a SHA-256")
if isinstance(self.seed, bool) or not isinstance(self.seed, int) or self.seed < 0:
raise RunArtifactError("seed must be a non-negative integer")
if not isinstance(self.model_revision, str) or not is_git_revision(self.model_revision):
raise RunArtifactError("model_revision must be an exact 40-hex revision")
if self.model_revision != QWEN35_2B_REVISION:
raise RunArtifactError("model_revision does not match pinned Qwen3.5-2B")
for name in (
"model_snapshot_sha256",
"environment_lock_sha256",
"evaluation_manifest_sha256",
"system_prompt_sha256",
):
if not is_sha256(str(getattr(self, name))):
raise RunArtifactError(f"{name} must be a SHA-256")
numeric_positive = (
"max_prompt_tokens",
"max_completion_tokens",
"total_context_tokens",
"per_device_train_batch_size",
"gradient_accumulation_steps",
"world_size",
"generations_per_prompt",
"checkpoint_interval",
"max_completion_tokens_per_run",
"lora_rank",
"lora_alpha",
)
for name in numeric_positive:
value = getattr(self, name)
if isinstance(value, bool) or not isinstance(value, int) or value <= 0:
raise RunArtifactError(f"{name} must be a positive integer")
if self.max_prompt_tokens + self.max_completion_tokens != self.total_context_tokens:
raise RunArtifactError("prompt + completion tokens must equal total context")
if (
isinstance(self.learning_rate, bool)
or not isinstance(self.learning_rate, int | float)
or not math.isfinite(self.learning_rate)
or self.learning_rate <= 0
):
raise RunArtifactError("learning_rate must be finite and positive")
if (
isinstance(self.lora_dropout, bool)
or not isinstance(self.lora_dropout, int | float)
or not math.isfinite(self.lora_dropout)
or not 0 <= self.lora_dropout < 1
):
raise RunArtifactError("lora_dropout must be finite and in [0,1)")
if self.lora_target_modules != ("all-linear",):
raise RunArtifactError("controlled runs require lora_target_modules=all-linear")
for name in (
"model_path",
"dataset_path",
"dataset_asset_root",
"output_dir",
"initial_checkpoint_path",
"environment_lock_path",
"evaluation_manifest_path",
):
value = getattr(self, name)
if (
not isinstance(value, str)
or not value.strip()
or value.strip()
in {
"required",
"set_by_common_oom_smoke",
"set_to_match_global_batch",
"TODO",
"TBD",
}
):
raise RunArtifactError(f"{name} is unresolved")
if self.trainer_kind == "sft":
if self.max_optimizer_steps != -1:
raise RunArtifactError("SFT requires max_optimizer_steps=-1 for one epoch")
if Path(self.initial_checkpoint_path).resolve() != Path(self.model_path).resolve():
raise RunArtifactError(
"SFT initial_checkpoint_path must equal model_path (base initialization)"
)
elif self.run_mode == "smoke" and self.max_optimizer_steps != 20:
raise RunArtifactError("controlled RL smoke requires exactly 20 optimizer steps")
elif self.run_mode == "main" and self.max_optimizer_steps != 5750:
raise RunArtifactError("controlled main RL runs require exactly 5,750 optimizer steps")
if self.run_mode == "main":
if (
not isinstance(self.compatibility_gate_path, str)
or not self.compatibility_gate_path.strip()
):
raise RunArtifactError("main RL runs require compatibility_gate_path")
elif self.compatibility_gate_path is not None:
raise RunArtifactError("SFT and smoke plans may not carry a compatibility gate")
if self.precision != "bf16":
raise RunArtifactError("Qwen3.5 controlled runs require bf16 precision")
if self.attention_implementation != "flash_attention_2":
raise RunArtifactError("Qwen3.5 controlled runs require flash_attention_2")
if not isinstance(self.use_vllm, bool):
raise RunArtifactError("use_vllm must be boolean")
if self.backend_entrypoint != "explicit_learning.training.backend:run":
raise RunArtifactError("controlled runs require explicit_learning.training.backend:run")
if (self.lora_rank, self.lora_alpha, self.lora_dropout) != (64, 128, 0.0):
raise RunArtifactError("controlled runs require LoRA r=64, alpha=128, dropout=0")
if self.learning_rate != 1e-6:
raise RunArtifactError("controlled runs require learning_rate=1e-6")
if (self.max_prompt_tokens, self.max_completion_tokens, self.total_context_tokens) != (
4096,
256,
4352,
):
raise RunArtifactError("controlled runs require a 4096/256/4352 token envelope")
expected_generations = 1 if self.trainer_kind == "sft" else 4
if self.generations_per_prompt != expected_generations:
raise RunArtifactError(
f"{self.trainer_kind} requires generations_per_prompt={expected_generations}"
)
if (
self.trainer_kind != "sft"
and self.global_prompt_batch_size % self.generations_per_prompt != 0
):
raise RunArtifactError(
"global prompt batch size must be divisible by generations_per_prompt"
)
token_cap = 6_000_000 if self.trainer_kind == "sft" else 24_000_000
if self.max_completion_tokens_per_run > token_cap:
raise RunArtifactError(
f"{self.trainer_kind} token budget exceeds protocol cap {token_cap}"
)
@property
def global_prompt_batch_size(self) -> int:
return self.per_device_train_batch_size * self.gradient_accumulation_steps * self.world_size
def to_dict(self) -> dict[str, Any]:
value = asdict(self)
value["lora_target_modules"] = list(self.lora_target_modules)
value["global_prompt_batch_size"] = self.global_prompt_batch_size
value["effective_completion_batch_size"] = self.global_prompt_batch_size
value["unique_prompt_groups_per_optimizer_step"] = (
self.global_prompt_batch_size
if self.trainer_kind == "sft"
else self.global_prompt_batch_size // self.generations_per_prompt
)
value["train_full_weights"] = False
value["base_model_resource"] = "qwen35_2b"
value["base_model_repo_id"] = "Qwen/Qwen3.5-2B"
value["training_contract_version"] = TRAINING_CONTRACT_VERSION
value["completion_only_sft_loss"] = self.trainer_kind == "sft"
value["assistant_only_sft_loss"] = False
value["reject_overlength_samples"] = True
# No-thinking protocol: the system prompt, SFT gold, structured rollout
# regex, and reward all share the same concise answer contract.
value["enable_thinking"] = False
value["structured_output_regex"] = (
None if self.trainer_kind == "sft" else ANSWER_SCHEMA_REGEX
)
value["optimizer"] = {
"name": "adamw_torch",
"learning_rate": self.learning_rate,
"weight_decay": 0.0,
"betas": [0.9, 0.999],
"eps": 1.0e-8,
}
value["scheduler"] = {"name": "cosine", "warmup_ratio": 0.03}
value["gradient_checkpointing"] = True
value["max_grad_norm"] = 1.0
value["num_train_epochs"] = 1
value["max_optimizer_steps"] = self.max_optimizer_steps
value["temperature"] = None if self.trainer_kind == "sft" else 1.0
value["top_p"] = None if self.trainer_kind == "sft" else 1.0
value["loss_type"] = (
"assistant_only_cross_entropy"
if self.trainer_kind == "sft"
else ("grpo" if self.trainer_kind == "papo" else "dr_grpo")
)
value["beta"] = (
None if self.trainer_kind == "sft" else (0.01 if self.trainer_kind == "papo" else 0.0)
)
value["freeze_vision_tower"] = False
value["freeze_multimodal_projector"] = False
value["lora_exclude_modules"] = ["embeddings", "lm_head"]
value["papo_config"] = (
{
"variant": "PAPO-G",
"mask_ratio": self.papo_mask_ratio,
"mask_type": "random",
"perception_loss_weight": self.papo_perception_loss_weight,
"der_loss_weight1": 0.0,
"der_loss_weight2": 0.0,
"require_gpu_contract_probe": self.run_mode == "smoke",
"adapter_version": PAPO_ADAPTER_VERSION,
"upstream_source_url": PAPO_UPSTREAM_SOURCE_URL,
"upstream_source_sha256": PAPO_UPSTREAM_SOURCE_SHA256,
}
if self.trainer_kind == "papo"
else None
)
value["evi_po_config"] = (
{
"adapter_version": EVI_PO_ADAPTER_VERSION,
"lambda_direction": self.evi_direction_loss_weight,
"lambda_evidence": self.evi_evidence_loss_weight,
"margin": self.evi_direction_margin,
"required_relationships": ["FULL", "CONTROL", "MISSING"],
"optional_relationships": ["SUBSTITUTE"],
"candidate_support": "group_answers_plus_unanswerable",
"evidence_supervision_relation": "FULL",
"evidence_sources": [
"executor_dependency_nodes_projected_through_node_map",
"source_annotation_explicit_mask",
"source_annotation_bboxes",
],
"attention_capture": "last_full_attention_layer_eager_forward_hook",
"require_gpu_contract_probe": (
self.run_mode == "smoke"
and (
self.evi_direction_loss_weight > 0.0 or self.evi_evidence_loss_weight > 0.0
)
),
"zero_weight_reduction": "exact_parent_grpo_path",
}
if self.trainer_kind == "evi_po"
else None
)
value["vllm_config"] = (
{
"mode": "colocate",
# Throughput-only knob: this is the fraction of GPU memory vLLM
# reserves for its KV cache during the rollout (generation) phase.
# Raising it lets more completions generate concurrently, which
# speeds up RL rollout — the dominant runtime cost — WITHOUT
# changing which tokens are sampled (seed + prompt + sampling
# params are unchanged), so completions, gradients, and the
# trained model are identical. The experimental batch sizes
# (per_device_train_batch_size, gradient_accumulation_steps,
# generations_per_prompt) are untouched. With enable_sleep_mode
# vLLM frees its KV cache for the batch-1 LoRA training step, so
# the rollout cap and the training step do not contend: a high
# cap enlarges only the rollout KV cache. 0.85 on 80GB H100s
# keeps ~67GB for vLLM weights+KV during rollout while the
# sleep-mode training step reclaims the KV portion. Lower if a
# colocate OOM appears (the smoke gate requires 0).
"gpu_memory_utilization": 0.85,
"max_model_length": self.total_context_tokens,
"tensor_parallel_size": 1,
"enable_sleep_mode": True,
"structured_outputs_regex": ANSWER_SCHEMA_REGEX,
}
if self.use_vllm
else None
)
value["token_budget_kind"] = (
"assistant_tokens" if self.trainer_kind == "sft" else "sampled_completion_tokens"
)
value["sampled_completion_token_budget_role"] = (
None if self.trainer_kind == "sft" else "hard_safety_ceiling_not_equality_target"
)
value["requires_forced_five_step_process_resume"] = self.run_mode == "smoke"
return value
def build_run_manifest(
config: RuntimeTrainingConfig,
*,
source_config_sha256: str,
dataset_manifest_sha256: str,
initial_checkpoint_sha256: str,
code_commit: str,
created_at: str,
launcher_command: tuple[str, ...],
) -> tuple[dict[str, Any], dict[str, Any]]:
"""Return frozen config and a manifest that says ``planned``, never trained."""
for name, value in (
("source_config_sha256", source_config_sha256),
("dataset_manifest_sha256", dataset_manifest_sha256),
("initial_checkpoint_sha256", initial_checkpoint_sha256),
):
if not isinstance(value, str) or not is_sha256(value):
raise RunArtifactError(f"{name} must be a SHA-256")
if not isinstance(code_commit, str) or not is_git_revision(code_commit):
raise RunArtifactError("code_commit must be an exact 40-hex revision")
if not isinstance(created_at, str) or not created_at:
raise RunArtifactError("created_at must be supplied")
try:
parsed_created_at = datetime.fromisoformat(created_at.replace("Z", "+00:00"))
except ValueError as exc:
raise RunArtifactError("created_at must be ISO-8601") from exc
if parsed_created_at.tzinfo is None:
raise RunArtifactError("created_at must include a timezone")
if not launcher_command or any(not str(part) for part in launcher_command):
raise RunArtifactError("launcher_command must be explicit and non-empty")
frozen = {
"schema_version": 1,
"source_config_sha256": source_config_sha256,
"runtime": config.to_dict(),
}
frozen_sha = canonical_json_hash(frozen)
manifest = {
"schema_version": 1,
"kind": "training_run_plan",
"status": "planned",
"trained": False,
"run_id": config.run_id,
"arm": config.arm,
"seed": config.seed,
"trainer_kind": config.trainer_kind,
"run_mode": config.run_mode,
"code_commit": code_commit,
"created_at": created_at,
"frozen_config_sha256": frozen_sha,
"dataset_manifest_sha256": dataset_manifest_sha256,
"initial_checkpoint_sha256": initial_checkpoint_sha256,
"model_snapshot_sha256": config.model_snapshot_sha256,
"environment_lock_sha256": config.environment_lock_sha256,
"evaluation_manifest_sha256": config.evaluation_manifest_sha256,
"system_prompt_sha256": config.system_prompt_sha256,
"comparison_slot_manifest_sha256": config.comparison_slot_manifest_sha256,
"launcher_command": list(launcher_command),
}
return frozen, manifest
def _write_once(path: Path, value: dict[str, Any]) -> None:
if path.exists():
try:
existing = json.loads(path.read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
raise RunArtifactError(f"existing frozen artifact is invalid JSON: {path}") from exc
if canonical_json(existing) != canonical_json(value):
raise RunArtifactError(f"refusing to overwrite drifted frozen artifact: {path}")
return
atomic_write_json(path, value)
def write_frozen_run(
run_dir: str | Path,
*,
frozen_config: dict[str, Any],
run_manifest: dict[str, Any],
) -> tuple[Path, Path]:
directory = Path(run_dir)
directory.mkdir(parents=True, exist_ok=True)
config_path = directory / "frozen-config.json"
manifest_path = directory / "run-manifest.json"
_write_once(config_path, frozen_config)
_write_once(manifest_path, run_manifest)
return config_path, manifest_path