Spaces:
Sleeping
Sleeping
File size: 16,421 Bytes
5850885 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 | """SQLDrift composed rubric (correctness, drift, speedup, gates, DBA tax).
Six child rubrics, one per reward component (:data:`REWARD_COMPONENT_KEYS`):
r_correct correctness vs ground-truth hash, gated on β₯ 1.2Γ speedup
r_drift bonus/penalty for (not) adapting to post-drift identifiers
r_speedup tanh-shaped speedup bonus, gated on r_correct > 0
r_step_tax base step tax plus bounded productive-action rebates
r_gatekeepers escalating tool-error / repeat-failing / no-op penalties
r_consult_dba DBA-oracle consult penalties (feature-flagged; 0 when off)
All child rubrics share a single ``ctx_provider`` that returns the private
:class:`engine.runtime.RuntimeEpisodeState`; this keeps the rubric
stateless relative to the environment and makes each component
individually unit-testable with a synthesized triple
``(RuntimeEpisodeState, SqlDriftAction, SqlDriftObservation)``.
"""
from __future__ import annotations
import math
import re
from collections.abc import Callable
from typing import TYPE_CHECKING
from openenv.core.rubrics import Rubric
from models import (
SqlDriftAction,
SqlDriftObservation,
SubmitRewriteResult,
ToolError,
ToolName,
)
if TYPE_CHECKING:
from engine.runtime import RuntimeEpisodeState
# Tunable thresholds pulled out to module level so tests and future
# curriculum code share a single source of truth.
SPEEDUP_MIN: float = 1.2
SPEEDUP_CAP_FOR_INFTY: float = 64.0
STEP_TAX: float = -0.03
STEP_REBATE_LIST_TABLES: float = 0.04
STEP_REBATE_DESCRIBE_TABLE: float = 0.06
STEP_REBATE_SAMPLE_ROWS: float = 0.05
STEP_REBATE_RUN_QUERY: float = 0.04
STEP_REBATE_EXPLAIN_QUERY: float = 0.04
STEP_REBATE_READ_CHANGELOG: float = 0.08
GATE_MALFORMED_TOOL_CALL: float = -0.3
GATE_CONSECUTIVE_TOOL_ERROR: float = -0.1
GATE_REPEAT_FAILING_QUERY: float = -0.1
GATE_BASELINE_VERBATIM: float = -0.2
_MAX_ESCALATION_STEPS: int = 3
CONSULT_ESCALATION: tuple[float, float, float] = (-0.1, -0.3, -0.8)
# =============================================================================
# Helpers
# =============================================================================
def canonicalize_sql(sql: str) -> str:
"""Whitespace/case/alias-insensitive canonical form.
Uses sqlglot's duckdb dialect round-trip so reorders/reformats agree;
falls back to a simple whitespace fold if sqlglot rejects the SQL
(e.g. during the baseline-verbatim check on an agent-submitted blob).
"""
try:
import sqlglot
expr = sqlglot.parse_one(sql, dialect="duckdb")
return expr.sql(dialect="duckdb", comments=False, normalize=True).strip().lower()
except Exception:
return " ".join(sql.lower().split())
_IDENT_RE = re.compile(r"\b([A-Za-z_][A-Za-z0-9_]*)\b")
def _extract_identifiers(sql: str) -> frozenset[str]:
"""Loose case-preserving identifier harvest.
Strips string literals before tokenizing so e.g. `WHERE x = 'ACTIVE'`
does not leak 'ACTIVE' into the identifier set.
"""
stripped = re.sub(r"'[^']*'", "", sql)
return frozenset(m.group(0) for m in _IDENT_RE.finditer(stripped))
def _extract_column_references(sql: str) -> frozenset[str]:
"""Column-reference identifiers via sqlglot AST (excludes alias labels).
For drift-adapt scoring, ``SELECT account_id AS user_id`` references
the new column and merely labels the output β the alias must not
count as a surviving pre-drift marker. Falls back to the regex
extractor on parse failure so malformed SQL still scores something.
"""
try:
import sqlglot
expr = sqlglot.parse_one(sql, dialect="duckdb")
except Exception:
return _extract_identifiers(sql)
if expr is None:
return _extract_identifiers(sql)
return frozenset(
n.name for n in expr.walk() if isinstance(n, sqlglot.exp.Column) and n.name
)
def _literals(sql: str) -> frozenset[str]:
"""All `'..'`-quoted string literals in `sql`."""
return frozenset(re.findall(r"'([^']*)'", sql))
_AGENT_MS_EPSILON: float = 1e-6
def effective_speedup(rt: RuntimeEpisodeState) -> float | None:
"""Compute ``effective_speedup`` from the runtime snapshot (speedup rubric).
Single source of truth for the speedup number used across the code
base β rubric scoring, the skill library's ``avg_speedup`` field,
and the training evaluator all route through here so divergent
definitions cannot drift apart.
Returns:
* ``None`` β no submission has happened yet. Callers that need a
numeric default (e.g. the rubric, which is only invoked
post-submission) should verify ``rt.submitted`` first.
* ``+β`` β drift has fired and the pre-drift baseline SQL no longer
executes against the post-drift schema; any correct submission is
definitionally "infinitely faster" than an unrunnable baseline.
* ``baseline_ms / max(agent_ms, Ξ΅)`` otherwise. A tiny ``Ξ΅`` clamp
guards against zero/negative timings from sub-microsecond queries
and treats them as "as fast as possible" (very large, finite
speedup) rather than silently collapsing the reward.
"""
if rt.submitted_runtime_ms is None:
return None
if rt.drift_fired and rt.baseline_postdrift_raises:
return math.inf
agent_ms = max(rt.submitted_runtime_ms, _AGENT_MS_EPSILON)
return rt.baseline_runtime_ms / agent_ms
def _speedup_for_reward(rt: RuntimeEpisodeState) -> float:
"""Rubric-facing speedup that never returns ``None``.
The rubric is only invoked once ``rt.submitted`` is True, so
:func:`effective_speedup` cannot return ``None`` from these call
sites; we assert that and coerce to ``0.0`` defensively if it ever
does (prevents a silent ``TypeError`` inside the reward math).
"""
val = effective_speedup(rt)
return 0.0 if val is None else val
def _is_terminal_submission(
action: SqlDriftAction,
observation: SqlDriftObservation,
rt: RuntimeEpisodeState,
) -> bool:
"""True iff this step is the submission step.
The env sets ``done=True`` on a successful submission and attaches a
:class:`SubmitRewriteResult`; we gate terminal rewards on both
signals so repeated rubric calls on an unchanged state don't
double-score.
"""
if not rt.submitted:
return False
if action.tool != ToolName.SUBMIT_REWRITE:
return False
tr = observation.tool_result
return isinstance(tr, SubmitRewriteResult)
def _gt_hash(rt: RuntimeEpisodeState) -> str | None:
if rt.drift_fired and rt.gt_result_hash_postdrift is not None:
return rt.gt_result_hash_postdrift
return rt.gt_result_hash_predrift
# =============================================================================
# Child rubrics
# =============================================================================
class _CtxChild(Rubric):
"""Base child rubric sharing the ctx provider."""
def __init__(self, ctx_provider: Callable[[], RuntimeEpisodeState]) -> None:
super().__init__()
object.__setattr__(self, "_ctx", ctx_provider)
def forward(
self,
action: SqlDriftAction,
observation: SqlDriftObservation,
) -> float:
raise NotImplementedError
class Correctness(_CtxChild):
"""Terminal-only correctness: +1.0 / +0.5 / -1.0 by hash and speedup."""
def forward(self, action: SqlDriftAction, observation: SqlDriftObservation) -> float:
rt = self._ctx()
if not _is_terminal_submission(action, observation, rt):
return 0.0
gt = _gt_hash(rt)
agent_hash = rt.submitted_result_hash
if gt is None or agent_hash is None:
return 0.0
if agent_hash != gt:
return -1.0
speedup = _speedup_for_reward(rt)
if speedup >= SPEEDUP_MIN:
return 1.0
return 0.5
class DriftAdapt(_CtxChild):
"""+0.5 for a correctly-adapted submission, -0.5 for a pre-drift-only
submission after drift fired.
Adaptation is detected against two scenario-declared identifier sets:
* ``postdrift_identifiers`` β identifiers/literals that only a
correct post-drift rewrite will introduce (e.g. ``account_id``
after a column rename, ``'ACTIVE'`` after an enum split).
* ``predrift_identifiers`` β identifiers/literals a submission that
ignored the drift would retain (e.g. ``user_id``, ``'active'``,
the ISO anchor strings under date-format drift).
A submission is considered "adapted" when it either surfaces a
post-drift marker *or* the scenario declares no distinctive
post-drift identifiers (e.g. date-format drift keeps the same
column name and only the literal shape changes) AND it does not
retain any pre-drift marker. The penalty fires only when the
submission still carries pre-drift markers AND produced the wrong
post-drift result β so a merely partial rewrite (neither pre-
nor post-flavoured) never earns a penalty it can't diagnose.
"""
def forward(self, action: SqlDriftAction, observation: SqlDriftObservation) -> float:
rt = self._ctx()
# Only drift scenarios participate.
if rt.gt_result_hash_postdrift is None and not rt.drift_fired:
return 0.0
if not _is_terminal_submission(action, observation, rt):
return 0.0
inst = getattr(rt, "instance", None)
post_ids: frozenset[str] = (
getattr(inst, "postdrift_identifiers", frozenset()) or frozenset()
)
pre_ids: frozenset[str] = getattr(inst, "predrift_identifiers", frozenset()) or frozenset()
agent_sql = rt.submitted_sql or ""
idents = _extract_column_references(agent_sql)
literals = _literals(agent_sql)
markers = idents | literals
uses_post = bool(post_ids & markers)
uses_pre = bool(pre_ids & markers)
# Treat "no distinctive post identifier" scenarios as
# satisfied by absence-of-pre (see class docstring).
adapted = (uses_post or not post_ids) and not uses_pre
agent_hash = rt.submitted_result_hash
gt_post = rt.gt_result_hash_postdrift
if rt.drift_fired and agent_hash == gt_post and adapted:
return 0.5
if rt.drift_fired and uses_pre and agent_hash != gt_post:
return -0.5
return 0.0
class Speedup(_CtxChild):
"""Terminal-only, gated on r_correct > 0: 0.3Β·tanh(log2(speedup)/3)."""
def forward(self, action: SqlDriftAction, observation: SqlDriftObservation) -> float:
rt = self._ctx()
if not _is_terminal_submission(action, observation, rt):
return 0.0
gt = _gt_hash(rt)
if gt is None or rt.submitted_result_hash != gt:
return 0.0
raw = _speedup_for_reward(rt)
if math.isinf(raw):
raw = SPEEDUP_CAP_FOR_INFTY
if raw <= 1.0:
return 0.0
return 0.3 * math.tanh(math.log2(raw) / 3.0)
class StepTax(_CtxChild):
"""Base step tax plus bounded rebates for productive exploration."""
def forward(self, action: SqlDriftAction, observation: SqlDriftObservation) -> float:
rt = self._ctx()
if _is_terminal_submission(action, observation, rt):
return 0.0
rebate = max(0.0, float(getattr(rt, "last_step_productive_rebate", 0.0)))
return STEP_TAX + rebate
class Gatekeepers(_CtxChild):
"""Sum of three independent penalties; repeats escalate up to a cap."""
def forward(self, action: SqlDriftAction, observation: SqlDriftObservation) -> float:
rt = self._ctx()
penalty = 0.0
# 1. Malformed / failed tool call β ToolError emitted this step.
if isinstance(observation.tool_result, ToolError):
penalty += GATE_MALFORMED_TOOL_CALL
streak = max(0, int(getattr(rt, "consecutive_tool_errors", 0)) - 1)
penalty += GATE_CONSECUTIVE_TOOL_ERROR * min(streak, _MAX_ESCALATION_STEPS)
# 2. Repeat failing query β env marks the flag on the runtime
# state immediately before invoking the rubric.
repeats = max(0, int(getattr(rt, "last_step_repeat_failing_query_count", 0)) - 1)
if repeats > 0:
penalty += GATE_REPEAT_FAILING_QUERY * min(repeats, _MAX_ESCALATION_STEPS)
# 3. Baseline-verbatim submission (Rev-3 gate β stacks with
# correctness's +0.5 partial to cap the no-op rewrite at +0.3).
if (
action.tool == ToolName.SUBMIT_REWRITE
and _is_terminal_submission(action, observation, rt)
and rt.submitted_sql_canonical == rt.baseline_sql_canonical
):
penalty += GATE_BASELINE_VERBATIM
return penalty
class ConsultDBA(_CtxChild):
"""Escalating penalties -0.1 / -0.3 / -0.8 per consult when the flag is on."""
def forward(self, action: SqlDriftAction, observation: SqlDriftObservation) -> float:
rt = self._ctx()
oracle_enabled = getattr(rt, "dba_oracle_enabled", False)
if not oracle_enabled:
return 0.0
if action.tool != ToolName.CONSULT_DBA:
return 0.0
# Count the consult THIS step by indexing into the escalation
# table using the pre-increment value (env increments on the same step).
tier = min(rt.consultations_used, len(CONSULT_ESCALATION))
if tier <= 0:
return CONSULT_ESCALATION[0]
return CONSULT_ESCALATION[tier - 1]
# =============================================================================
# Composite
# =============================================================================
class SqlDriftRubric(Rubric):
"""Composite rubric: sum of six children.
Registration as attributes auto-enrolls them in
:meth:`Rubric.named_rubrics` so training loops can introspect
per-component scores.
"""
def __init__(self, ctx_provider: Callable[[], RuntimeEpisodeState]) -> None:
super().__init__()
# NOTE: order matters β correctness must populate last_score before
# speedup reads it via the shared ctx_provider (both are pure
# functions of the runtime state, so identical output β but the
# explicit ordering documents the intent).
self.correctness = Correctness(ctx_provider)
self.drift_adapt = DriftAdapt(ctx_provider)
self.speedup = Speedup(ctx_provider)
self.step_tax = StepTax(ctx_provider)
self.gatekeepers = Gatekeepers(ctx_provider)
self.consult_dba = ConsultDBA(ctx_provider)
def forward(self, action: SqlDriftAction, observation: SqlDriftObservation) -> float:
total = (
self.correctness(action, observation)
+ self.drift_adapt(action, observation)
+ self.speedup(action, observation)
+ self.step_tax(action, observation)
+ self.gatekeepers(action, observation)
+ self.consult_dba(action, observation)
)
return total
def component_scores(self) -> dict[str, float]:
"""Return the most-recent per-component scores, keyed for W&B.
Keys match :data:`models.REWARD_COMPONENT_KEYS` so the observation
envelope and the demo plots agree on a stable schema.
"""
return {
"r_correct": float(self.correctness.last_score or 0.0),
"r_drift": float(self.drift_adapt.last_score or 0.0),
"r_speedup": float(self.speedup.last_score or 0.0),
"r_step_tax": float(self.step_tax.last_score or 0.0),
"r_gatekeepers": float(self.gatekeepers.last_score or 0.0),
"r_consult_dba": float(self.consult_dba.last_score or 0.0),
}
__all__ = [
"CONSULT_ESCALATION",
"ConsultDBA",
"Correctness",
"DriftAdapt",
"GATE_BASELINE_VERBATIM",
"GATE_CONSECUTIVE_TOOL_ERROR",
"GATE_MALFORMED_TOOL_CALL",
"GATE_REPEAT_FAILING_QUERY",
"Gatekeepers",
"SPEEDUP_CAP_FOR_INFTY",
"SPEEDUP_MIN",
"STEP_REBATE_DESCRIBE_TABLE",
"STEP_REBATE_EXPLAIN_QUERY",
"STEP_REBATE_LIST_TABLES",
"STEP_REBATE_READ_CHANGELOG",
"STEP_REBATE_RUN_QUERY",
"STEP_REBATE_SAMPLE_ROWS",
"STEP_TAX",
"Speedup",
"SqlDriftRubric",
"StepTax",
"canonicalize_sql",
"effective_speedup",
]
|