Spaces:
Sleeping
Sleeping
File size: 11,944 Bytes
eb1ebe6 8fa7af1 43f41de 8fa7af1 43f41de 8fa7af1 43f41de eb1ebe6 43f41de eb1ebe6 8fa7af1 5869d56 8fa7af1 5869d56 eb1ebe6 43f41de 8fa7af1 43f41de 8fa7af1 43f41de 8fa7af1 43f41de eb1ebe6 43f41de eb1ebe6 8fa7af1 43f41de eb1ebe6 43f41de eb1ebe6 43f41de 8fa7af1 43f41de 8fa7af1 43f41de eb1ebe6 43f41de 8fa7af1 43f41de eb1ebe6 43f41de 8fa7af1 43f41de eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 43f41de eb1ebe6 8fa7af1 eb1ebe6 43f41de eb1ebe6 8fa7af1 eb1ebe6 43f41de eb1ebe6 8fa7af1 eb1ebe6 43f41de eb1ebe6 8fa7af1 eb1ebe6 43f41de 8fa7af1 43f41de 8fa7af1 eb1ebe6 8fa7af1 eb1ebe6 8fa7af1 eb1ebe6 43f41de 8fa7af1 eb1ebe6 43f41de 8fa7af1 43f41de 8fa7af1 43f41de eb1ebe6 8fa7af1 eb1ebe6 8fa7af1 eb1ebe6 43f41de 8fa7af1 43f41de 8fa7af1 43f41de 8fa7af1 43f41de 8fa7af1 43f41de 8fa7af1 43f41de 5869d56 43f41de 8fa7af1 43f41de 8fa7af1 43f41de 8fa7af1 | 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 | """Reward components for the generation phase.
After exploration, the agent generates marimo/manim code. Rewards measure
validity, task alignment, artifact structure, and research usage.
Scoring model:
quality = weighted sum of (validity, task alignment, structure, research usage)
total = quality × gate
Gates (multiplicative):
- code doesn't parse → total = 0
- static check fails → total = quality × small static-fail multiplier
- code doesn't run → total = quality × execution-fail multiplier
- code runs → total = quality × 1.0
"""
from __future__ import annotations
import re
from typing import TYPE_CHECKING
from .sandbox import ast_parses, check_marimo, extract_scene_class
try:
from ..constants import MAX_REPAIR_REWARD, clamp_action_reward
except ImportError: # pragma: no cover - supports direct test execution
from constants import MAX_REPAIR_REWARD, clamp_action_reward
if TYPE_CHECKING:
from ..task_bank import Task
# ---------------------------------------------------------------------------
# Component weights
# ---------------------------------------------------------------------------
_WEIGHTS = {
"validity": 0.15,
"task_alignment": 0.30,
"structure": 0.30,
"research_usage": 0.25,
}
GATE_STATIC_FAIL = 0.12
GATE_RUNS_FAIL = 0.30 # quality multiplier when static checks pass but execution fails
_STOPWORDS = {
"about", "after", "again", "against", "also", "because", "before", "being",
"between", "class", "code", "construct", "could", "from", "have", "into",
"like", "make", "more", "most", "only", "self", "show", "step", "than",
"that", "their", "then", "there", "these", "this", "through", "using",
"value", "where", "with", "would",
}
# ---------------------------------------------------------------------------
# Individual scorers
# ---------------------------------------------------------------------------
def keyword_coverage(code: str, keywords_csv: str) -> float:
"""Fraction of task keywords mentioned in the code (case-insensitive)."""
if not keywords_csv:
return 0.0
keywords = [k.strip().lower() for k in keywords_csv.split(",") if k.strip()]
if not keywords:
return 0.0
code_lower = code.lower()
return sum(1 for kw in keywords if kw in code_lower) / len(keywords)
def format_match(chosen_format: str, task: Task) -> float:
"""1.0 if format matches the task's preferred format, else 0.3.
If the task has no preferred format (None), any choice scores 1.0.
"""
if task.preferred_format is None:
return 1.0
return 1.0 if chosen_format == task.preferred_format else 0.3
def marimo_structure(
code: str,
task: Task,
static_check_passed: bool | None = None,
error_codes: list[str] | None = None,
) -> float:
"""Score structural quality of a marimo notebook (0-1).
Additive scoring for good patterns, penalties from ``marimo check``
for breaking violations (duplicate defs, cycles, etc.).
"""
score = 0.0
# Positive signals
if "import marimo" in code or "from marimo" in code:
score += 0.2
if "marimo.App" in code or "mo.App" in code:
score += 0.1
cell_count = code.count("@app.cell")
if cell_count >= 3:
score += 0.2
elif cell_count >= 1:
score += 0.1
ui_patterns = [
"mo.md(",
"mo.Html",
"mo.accordion",
"mo.callout",
"mo.hstack(",
"mo.vstack(",
"mo.ui.slider",
"mo.ui.dropdown",
"mo.ui.table",
"mo.ui.dataframe",
]
score += min(0.22, sum(0.06 for p in ui_patterns if p in code))
reactive_plot_patterns = [
"mo.ui.matplotlib(",
"mo.ui.plotly(",
"mo.ui.altair_chart(",
]
raw_plot_patterns = [
"plt.",
"matplotlib.pyplot",
"px.",
"plotly.",
"alt.Chart",
]
if "mo.ui.matplotlib(plt.gca())" in code:
score += 0.24 if task.data_available else 0.16
elif any(p in code for p in reactive_plot_patterns):
score += 0.18 if task.data_available else 0.10
elif any(p in code for p in raw_plot_patterns):
score += 0.08 if task.data_available else 0.03
score -= 0.08
if "plt.tight_layout(" in code:
score -= 0.12
if "np.math." in code:
score -= 0.15
tier_thresholds = {"advanced": 6, "intermediate": 4, "beginner": 2}
if cell_count >= tier_thresholds.get(task.tier, 2):
score += 0.1
# Marimo check: penalize breaking violations, bonus for clean code
if static_check_passed is None:
passed, _, violations = check_marimo(code)
else:
passed = static_check_passed
violations = error_codes or []
if passed:
score += 0.1
else:
penalty = {
"MB002": 0.35,
"MB003": 0.4,
"MB005": 0.25,
"MB001": 0.3,
"MB004": 0.2,
}
for v in violations:
score -= penalty.get(v, 0.15)
return max(0.0, min(1.0, score))
def manim_structure(code: str, task: Task) -> float:
"""Score structural quality of a manim scene (0-1)."""
from .sandbox import extract_scene_class
score = 0.0
if "from manim" in code or "import manim" in code:
score += 0.2
if extract_scene_class(code) is not None:
score += 0.2
if "def construct" in code:
score += 0.1
anim_patterns = [
"self.play(", "self.wait(", "Create(", "FadeIn(", "FadeOut(",
"Transform(", "Write(", "MoveToTarget", "Indicate(",
"ReplacementTransform(",
]
anim_hits = sum(1 for p in anim_patterns if p in code)
score += min(0.3, anim_hits * 0.05)
math_patterns = ["MathTex(", "Tex(", "Axes(", "NumberPlane(", "Graph("]
if any(p in code for p in math_patterns):
score += 0.1
tier_thresholds = {"advanced": 6, "intermediate": 4, "beginner": 2}
if anim_hits >= tier_thresholds.get(task.tier, 2):
score += 0.1
return min(1.0, score)
def narration_score(narration: str, fmt: str) -> float:
"""Score narration quality. Only relevant for manim format."""
if fmt != "manim":
return 1.0
if not narration or not narration.strip():
return 0.0
words = narration.split()
score = 0.0
if len(words) >= 30:
score += 0.4
elif len(words) >= 10:
score += 0.2
scene_markers = ["scene", "step", "first", "next", "then", "finally", "now"]
score += min(0.3, sum(0.1 for m in scene_markers if m in narration.lower()))
if len(words) >= 50:
score += 0.3
elif len(words) >= 20:
score += 0.15
return min(1.0, score)
def context_usage(code: str, accumulated_context: list[str]) -> float:
"""Score whether the generated code incorporates research findings (0-1)."""
if not accumulated_context:
return 0.0
context_words: set[str] = set()
for ctx in accumulated_context:
context_words.update(_tokens(ctx))
if not context_words:
return 0.0
code_words = set(_tokens(code))
overlap = code_words & context_words
if not overlap:
return 0.0
# Do not reward broad generic overlap too heavily; a few meaningful terms
# should help, but strong usage needs a substantial slice of the context.
target = min(max(len(context_words), 1), 24)
return min(1.0, len(overlap) / target * 2.5)
# ---------------------------------------------------------------------------
# Main reward function
# ---------------------------------------------------------------------------
def compute_generate_reward(
code: str,
fmt: str,
narration: str,
task: Task,
exec_success: bool,
accumulated_context: list[str],
static_check_passed: bool | None = None,
error_codes: list[str] | None = None,
) -> tuple[float, dict]:
"""Compute the generation-phase reward. Returns (total, components).
``python_parse_valid``, ``static_check_passed``, and ``code_runs`` act as
gates. ``code_valid`` means the artifact is valid for its target format,
not merely that the Python AST parses.
"""
parse_valid = ast_parses(code)
c_parse = 1.0 if parse_valid else 0.0
if static_check_passed is None:
static_check_passed = _infer_static_check(code, fmt, parse_valid)
c_static = 1.0 if parse_valid and static_check_passed else 0.0
c_runs = 1.0 if exec_success else 0.0
c_coverage = keyword_coverage(code, task.keywords)
c_format = format_match(fmt, task)
if fmt == "marimo":
c_struct = marimo_structure(code, task, static_check_passed, error_codes)
else:
scene_structure = manim_structure(code, task)
c_struct = 0.75 * scene_structure + 0.25 * narration_score(narration, fmt)
c_ctx = context_usage(code, accumulated_context)
c_validity = _validity_score(c_parse, c_static, c_runs)
c_alignment = 0.75 * c_coverage + 0.25 * c_format
quality = (
_WEIGHTS["validity"] * c_validity
+ _WEIGHTS["task_alignment"] * c_alignment
+ _WEIGHTS["structure"] * c_struct
+ _WEIGHTS["research_usage"] * c_ctx
)
# Apply gates
if c_parse == 0.0:
total = 0.0
elif c_static == 0.0:
total = quality * _static_fail_multiplier(error_codes or [])
elif c_runs == 0.0:
total = quality * GATE_RUNS_FAIL
else:
total = quality
components = {
"validity": round(c_validity, 3),
"task_alignment": round(c_alignment, 3),
"structure": round(c_struct, 3),
"research_usage": round(c_ctx, 3),
"generate_total": round(total, 4),
}
return total, components
def _infer_static_check(code: str, fmt: str, parse_valid: bool) -> bool:
if not parse_valid:
return False
if fmt == "marimo":
passed, _, _ = check_marimo(code)
return passed
if fmt == "manim":
return extract_scene_class(code) is not None
return False
def _static_fail_multiplier(error_codes: list[str]) -> float:
"""Keep parseable but structurally invalid artifacts from scoring high."""
if any(code.startswith("MB") for code in error_codes):
return GATE_STATIC_FAIL
return min(GATE_RUNS_FAIL, GATE_STATIC_FAIL * 1.5)
def _validity_score(
parse_valid: float,
static_check_passed: float,
code_runs: float,
) -> float:
if parse_valid == 0.0:
return 0.0
if static_check_passed == 0.0:
return 0.35
if code_runs == 0.0:
return 0.70
return 1.0
def adjust_repair_reward(
base_reward: float,
*,
repair_success: bool,
previous_error_codes: list[str],
new_error_codes: list[str],
previous_code: str,
repaired_code: str,
) -> tuple[float, dict]:
"""Discount repaired code but reward fixing the specific prior failure."""
changed = _fingerprint(previous_code) != _fingerprint(repaired_code)
fixed_prior = bool(previous_error_codes) and not (
set(previous_error_codes) & set(new_error_codes)
)
if repair_success:
reward = base_reward * 0.60
reward += 0.08 if fixed_prior else 0.0
reward += 0.04 if changed else 0.0
else:
reward = base_reward * 0.25
reward += 0.04 if fixed_prior else 0.0
if not changed:
reward -= 0.15
reward = min(MAX_REPAIR_REWARD, clamp_action_reward(reward))
return reward, {
"repair_success": 1.0 if repair_success else 0.0,
"fixed_prior_errors": 1.0 if fixed_prior else 0.0,
"changed_code": 1.0 if changed else 0.0,
"repair_total": round(reward, 4),
}
def _tokens(text: str) -> list[str]:
return [
w
for w in re.findall(r"\w+", text.lower())
if len(w) > 3 and w not in _STOPWORDS
]
def _fingerprint(code: str) -> str:
return re.sub(r"\s+", "", code)
|