File size: 8,923 Bytes
e2f8b29 206438f e2f8b29 206438f e2f8b29 206438f e2f8b29 206438f e2f8b29 206438f e2f8b29 206438f e2f8b29 206438f e2f8b29 56b42b0 e2f8b29 56b42b0 e2f8b29 56b42b0 e2f8b29 56b42b0 e2f8b29 56b42b0 e2f8b29 56b42b0 e2f8b29 56b42b0 e2f8b29 56b42b0 e2f8b29 206438f e2f8b29 56b42b0 02e58fe e2f8b29 02e58fe e2f8b29 56b42b0 02e58fe 56b42b0 02e58fe e2f8b29 02e58fe e2f8b29 02e58fe e2f8b29 02e58fe e2f8b29 56b42b0 e2f8b29 56b42b0 e2f8b29 0b9b77b 206438f 02e58fe 0b9b77b 02e58fe 0b9b77b 02e58fe 0b9b77b 02e58fe 0b9b77b 02e58fe 0b9b77b 02e58fe 0b9b77b e2f8b29 0b9b77b e2f8b29 f1b57dc e2f8b29 f1b57dc | 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 | """Per-task grader functions — returns normalized 0.0-1.0 score at episode end.
Separate from reward_engine.py. Evaluates EpisodeState holistically.
NOT a sum of step rewards.
"""
from __future__ import annotations
import torch # noqa: F401
from ml_training_debugger.models import EpisodeState
from ml_training_debugger.scenarios import ScenarioParams
FIX_ACTIONS = frozenset(
{
"modify_config",
"add_callback",
"replace_optimizer",
"patch_data_loader",
"fix_model_mode",
"fix_code",
}
)
def _has_action(state: EpisodeState, action_type: str) -> bool:
return action_type in state.actions_taken
def _correct_diagnosis(state: EpisodeState, scenario: ScenarioParams) -> bool:
if not state.diagnosis_submitted:
return False
# Find the diagnosis from actions_taken metadata
# We store "mark_diagnosed:<diagnosis>" in actions_taken
for action_str in reversed(state.actions_taken):
if action_str.startswith("mark_diagnosed:"):
submitted = action_str.split(":", 1)[1]
return submitted == scenario.root_cause.value
return False
def _submitted_diagnosis(state: EpisodeState) -> str | None:
for action_str in reversed(state.actions_taken):
if action_str.startswith("mark_diagnosed:"):
return action_str.split(":", 1)[1]
return None
def grade_task_001(state: EpisodeState, scenario: ScenarioParams) -> float:
"""Grade Task 1 — Exploding Gradients (easy)."""
score = 0.0
# +0.05 for inspect_gradients
if state.gradients_inspected:
score += 0.05
# +0.20 for correct fix (modify_config with LR reduction)
if _has_action(state, "modify_config"):
score += 0.20
# +0.35 for restart with convergence
if state.restart_after_fix:
score += 0.35
# +0.40 for correct diagnosis
if _correct_diagnosis(state, scenario):
score += 0.40
return min(1.0, max(0.0, score))
def grade_task_002(state: EpisodeState, scenario: ScenarioParams) -> float:
"""Grade Task 2 — Vanishing Gradients (easy)."""
score = 0.0
if state.gradients_inspected:
score += 0.05
if _has_action(state, "modify_config"):
score += 0.20
if state.restart_after_fix:
score += 0.35
if _correct_diagnosis(state, scenario):
score += 0.40
return min(1.0, max(0.0, score))
def grade_task_003(state: EpisodeState, scenario: ScenarioParams) -> float:
"""Grade Task 3 — Silent Data Leakage (medium)."""
score = 0.0
# +0.05 for inspect_data_batch
if state.data_inspected:
score += 0.05
# +0.30 for patch_data_loader
if _has_action(state, "patch_data_loader"):
score += 0.30
# +0.30 for restart with convergence (val accuracy normalizes)
if state.restart_after_fix:
score += 0.30
# +0.35 for correct diagnosis
if _correct_diagnosis(state, scenario):
score += 0.35
return min(1.0, max(0.0, score))
def grade_task_004(state: EpisodeState, scenario: ScenarioParams) -> float:
"""Grade Task 4 — Overfitting (medium)."""
score = 0.0
if state.data_inspected:
score += 0.05
if _has_action(state, "modify_config") or _has_action(state, "add_callback"):
score += 0.25
if state.restart_after_fix:
score += 0.30
if _correct_diagnosis(state, scenario):
score += 0.40
return min(1.0, max(0.0, score))
def grade_task_005(state: EpisodeState, scenario: ScenarioParams) -> float:
"""Grade Task 5 — BatchNorm Eval Mode (hard).
Hard task requires thorough investigation for full credit.
Context-gated penalty: -0.20 if add_callback after gradients_were_normal.
Wrong-fix penalty: -0.10 if modify_config used (LR isn't the problem).
Full credit requires inspecting gradients, weights, data, AND modes.
"""
score = 0.0
# Investigation credits (+0.05 each, max +0.15 for 3 types)
if state.gradients_inspected:
score += 0.05
if state.model_modes_inspected:
score += 0.05
if state.data_inspected:
score += 0.05
# Red herring penalty: -0.20 for add_callback after normal gradients
if (
_has_action(state, "add_callback")
and state.gradients_inspected
and state.gradients_were_normal
):
score -= 0.20
# Wrong-fix penalty: -0.10 for modify_config (LR isn't the issue here)
if _has_action(state, "modify_config"):
score -= 0.10
# Fix credit scaled by investigation thoroughness
# Full credit requires weight inspection (rules out weight-related causes)
if _has_action(state, "fix_model_mode"):
if state.model_weights_inspected and state.data_inspected:
score += 0.25 # Thorough: checked weights AND data
elif state.model_weights_inspected or state.data_inspected:
score += 0.15 # Partial thoroughness
else:
score += 0.08 # Quick fix without ruling out alternatives
# Restart credit scaled similarly
if state.restart_after_fix:
if state.model_weights_inspected:
score += 0.20 # Full restart credit
else:
score += 0.10 # Partial credit
# +0.40 for correct diagnosis
if _correct_diagnosis(state, scenario):
score += 0.40
return min(1.0, max(0.0, score))
def grade_task_006(state: EpisodeState, scenario: ScenarioParams) -> float:
"""Grade Task 6 — PyTorch Code Bug (hard).
Diagnosis must ALWAYS be 'code_bug' regardless of bug variant.
Hard task rewards thorough investigation before fixing.
Full credit requires ruling out non-code causes via weight inspection.
"""
score = 0.0
# Investigation credits (+0.05 each, up to +0.25 for all 5 types)
if state.code_inspected:
score += 0.05
if state.gradients_inspected:
score += 0.05
if state.model_modes_inspected:
score += 0.05
if state.model_weights_inspected:
score += 0.05
if state.data_inspected:
score += 0.05
# Code fix credit scaled by investigation thoroughness
if _has_action(state, "fix_code") and state.fix_action_taken:
if state.model_weights_inspected:
score += 0.15 # Thorough: ruled out weight-related causes
else:
score += 0.08 # Quick fix without full investigation
# Restart credit scaled by thoroughness
if state.restart_after_fix:
if state.model_weights_inspected:
score += 0.15 # Full restart credit
else:
score += 0.08 # Partial credit
# +0.45 for correct diagnosis (must be code_bug)
if _correct_diagnosis(state, scenario):
score += 0.45
return min(1.0, max(0.0, score))
def grade_task_007(state: EpisodeState, scenario: ScenarioParams) -> float:
"""Grade Task 7 — LR Scheduler Misconfigured (hard).
Requires thorough investigation: agents must inspect weights to rule out
weight-related issues before concluding scheduler is the root cause.
Penalizes wrong fixes (e.g. patch_data_loader when data is fine).
"""
score = 0.0
# Investigation credits (+0.05 each, up to +0.20 for all 4 types)
if state.gradients_inspected:
score += 0.05
if state.data_inspected:
score += 0.05
if state.model_weights_inspected:
score += 0.05
if state.model_modes_inspected:
score += 0.05
# Fix credit scaled by investigation thoroughness
if _has_action(state, "modify_config"):
if state.model_weights_inspected:
score += 0.20 # Thorough: ruled out weight issues
else:
score += 0.12 # Partial: didn't check weights
# Restart credit scaled by thoroughness
if state.restart_after_fix:
if state.model_weights_inspected:
score += 0.20 # Full restart credit
else:
score += 0.12 # Partial credit
# Diagnosis
if _correct_diagnosis(state, scenario):
score += 0.40
# Wrong-fix penalty: patch_data_loader when data is clean
if _has_action(state, "patch_data_loader"):
score -= 0.10
return min(1.0, max(0.0, score))
# Registry mapping task IDs to grader functions
GRADERS = {
"task_001": grade_task_001,
"task_002": grade_task_002,
"task_003": grade_task_003,
"task_004": grade_task_004,
"task_005": grade_task_005,
"task_006": grade_task_006,
"task_007": grade_task_007,
}
def grade_episode(task_id: str, state: EpisodeState, scenario: ScenarioParams) -> float:
"""Grade a completed episode. Returns score in (0.0, 1.0) exclusive."""
grader = GRADERS.get(task_id)
if grader is None:
return 0.01
score = grader(state, scenario)
# Clamp to strictly between 0 and 1 (evaluator rejects exact 0.0 and 1.0)
return max(0.01, min(0.99, score))
|