File size: 9,527 Bytes
df455fc ba2cf79 df455fc ba2cf79 c6bc2db df455fc ba2cf79 df455fc c6bc2db df455fc c6bc2db df455fc | 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 | from __future__ import annotations
from typing import Any
from .metrics import (
compute_aesthetics_scores,
compute_presentation_diagnostics,
compute_slide_staleness_penalty,
mean_scores_by_dimension,
redundancy_score,
score_checklist_item,
score_generic_slide_checklist_items,
score_slide_checklist_item,
)
from ..reward_metrics import (
clamp_reward,
compute_overlap_ratio,
is_blank_or_title_only,
)
from ..reward_models import (
IntermediateSlideRewardResult,
PresentBenchEvalSpec,
PresentBenchScoreResult,
ExtractedPresentation,
ExtractedSlide,
TaskSpec,
)
def score_presentbench(
task_spec: TaskSpec,
presentation_extraction: ExtractedPresentation,
eval_spec: PresentBenchEvalSpec,
*,
aesthetics_service: Any | None = None,
) -> PresentBenchScoreResult:
diagnostics = compute_presentation_diagnostics(presentation_extraction, task_spec)
checklist_results = [
score_checklist_item(item, presentation_extraction, task_spec, diagnostics)
for item in eval_spec.checklist
]
dimension_scores = mean_scores_by_dimension(checklist_results)
aesthetics_scores = (
aesthetics_service.score_presentation(presentation_extraction)
if aesthetics_service and hasattr(aesthetics_service, "score_presentation")
else compute_aesthetics_scores(presentation_extraction)
)
s_fund = dimension_scores.get("fundamentals", 0.0)
s_visual = dimension_scores.get("visual_layout", 0.0)
s_complete = dimension_scores.get("completeness", 0.0)
s_correct = dimension_scores.get("correctness", 0.0)
s_fidelity = dimension_scores.get("fidelity", 0.0)
dimension_weights = eval_spec.scoring_config["dimension_weights"]
r_pb = (
dimension_weights["fundamentals"] * s_fund
+ dimension_weights["visual_layout"] * s_visual
+ dimension_weights["completeness"] * s_complete
+ dimension_weights["correctness"] * s_correct
+ dimension_weights["fidelity"] * s_fidelity
)
c_fidelity_critical = (
0.5
if any(
r["dimension"] == "fidelity" and r["verdict"] == "no"
for r in checklist_results
)
else 1.0
)
blankness_threshold = eval_spec.scoring_config["hard_caps"][
"blank_title_only_ratio_threshold"
]
c_blankness = (
0.6 if diagnostics["blank_title_only_ratio"] > blankness_threshold else 1.0
)
c_hard = min(c_fidelity_critical, c_blankness)
penalty_config = eval_spec.scoring_config["soft_penalties"]
soft_penalties = {
"slide_count_violation": penalty_config["slide_count_violation"]
* diagnostics["slide_count_violation"],
"overlap": penalty_config["overlap"] * diagnostics["max_overlap_ratio"],
"tiny_text": penalty_config["tiny_text"]
* (
1.0
if diagnostics["min_font_size_pt"] is not None
and diagnostics["min_font_size_pt"] < 10
else 0.0
),
"staleness": penalty_config["staleness"]
* diagnostics["mean_staleness_penalty"],
}
p_soft = sum(soft_penalties.values())
aesthetic_weight = eval_spec.scoring_config.get("aesthetic_weight", 0.15)
reward_total = clamp_reward(
c_hard
* (
(1.0 - aesthetic_weight) * r_pb
+ aesthetic_weight * aesthetics_scores["aesthetic"]
)
- p_soft
)
return PresentBenchScoreResult(
reward_total=reward_total,
reward_breakdown={
"R_pb": reward_total,
"S_fundamentals": s_fund,
"S_visual_layout": s_visual,
"S_completeness": s_complete,
"S_correctness": s_correct,
"S_fidelity": s_fidelity,
"S_aesthetic": aesthetics_scores["aesthetic"],
"S_harmony": aesthetics_scores["harmony"],
"S_engagement": aesthetics_scores["engagement"],
"S_usability": aesthetics_scores["usability"],
"S_rhythm": aesthetics_scores["rhythm"],
},
hard_caps={
"C_fidelity_critical": c_fidelity_critical,
"C_blankness": c_blankness,
"C_hard": c_hard,
},
soft_penalties=soft_penalties,
checklist_results=checklist_results,
aesthetics_results=aesthetics_scores,
metadata={
"task_id": eval_spec.task_spec.task_id,
"item_count": len(eval_spec.checklist),
"slide_count": presentation_extraction.slide_count,
"spec_hash": eval_spec.spec_hash,
},
)
def score_presentbench_slide(
task_spec: TaskSpec,
eval_spec: PresentBenchEvalSpec,
slide_index: int,
slide_extraction: ExtractedSlide,
*,
previous_slide_extractions: list[ExtractedSlide] | None = None,
use_mllm: bool = False,
mode: str = "eval",
) -> IntermediateSlideRewardResult:
target_slides = {
slide.slide_index: slide for slide in task_spec.required_slides or []
}
target_slide = target_slides.get(slide_index)
target_mode = "required_slide" if target_slide is not None else "fallback_generic"
if target_slide is None:
checklist_results = score_generic_slide_checklist_items(
slide_extraction, task_spec
)
else:
slide_checklist = eval_spec.slide_checklists.get(slide_index, [])
checklist_results = [
score_slide_checklist_item(
item,
slide_extraction,
target_slide.slide_role,
target_slide.title_hint,
target_slide.required_shape_kinds,
task_spec,
)
for item in slide_checklist
]
dimension_scores = mean_scores_by_dimension(checklist_results)
s_prompt_alignment = dimension_scores.get("prompt_alignment", 0.0)
s_local_completeness = dimension_scores.get("local_completeness", 0.0)
s_local_correctness = dimension_scores.get("local_correctness", 0.0)
s_local_fidelity = dimension_scores.get("local_fidelity", 0.0)
s_local_usability = dimension_scores.get("local_usability", 0.0)
c_slide_fidelity_critical = (
0.5
if (
target_slide is not None
and s_local_fidelity < 1.0
and target_slide.required_exact_values
)
else 1.0
)
c_slide_blankness = (
0.4
if is_blank_or_title_only(slide_extraction)
and (
(target_slide is not None and target_slide.required_points)
or target_mode == "fallback_generic"
)
else 1.0
)
c_slide_hard = min(c_slide_fidelity_critical, c_slide_blankness)
penalty_config = eval_spec.scoring_config["soft_penalties"]
overlap = compute_overlap_ratio(slide_extraction)
min_font = slide_extraction.text_metrics.get("min_font_size_pt")
redundancy = redundancy_score(slide_extraction, previous_slide_extractions)
staleness = compute_slide_staleness_penalty(
slide_extraction,
role=target_slide.slide_role if target_slide is not None else None,
)
wrong_slot_behavior = (
0.0 if target_mode == "fallback_generic" or s_prompt_alignment >= 1.0 else 1.0
)
soft_penalties = {
"redundancy": penalty_config["redundancy"] * redundancy,
"wrong_slot_behavior": penalty_config["wrong_slot_behavior"]
* wrong_slot_behavior,
"tiny_text": penalty_config["tiny_text"]
* (1.0 if min_font is not None and min_font < 10 else 0.0),
"overlap": penalty_config["overlap"] * overlap,
"staleness": penalty_config["staleness"] * float(staleness["penalty"]),
}
p_slide_soft = sum(soft_penalties.values())
reward_total = clamp_reward(
c_slide_hard
* (
0.35 * s_prompt_alignment
+ 0.20 * s_local_completeness
+ 0.15 * s_local_correctness
+ 0.20 * s_local_fidelity
+ 0.10 * s_local_usability
)
- p_slide_soft
)
return IntermediateSlideRewardResult(
slide_index=slide_index,
reward_total=reward_total,
reward_breakdown={
"R_slide": reward_total,
"S_prompt_alignment": s_prompt_alignment,
"S_local_completeness": s_local_completeness,
"S_local_correctness": s_local_correctness,
"S_local_fidelity": s_local_fidelity,
"S_local_usability": s_local_usability,
},
hard_caps={
"C_slide_fidelity_critical": c_slide_fidelity_critical,
"C_slide_blankness": c_slide_blankness,
"C_slide_hard": c_slide_hard,
},
soft_penalties=soft_penalties,
checklist_results=checklist_results,
metadata={
"slide_index": slide_index,
"slide_id": slide_extraction.slide_id,
"target_mode": target_mode,
"target_slide_role": target_slide.slide_role if target_slide else None,
"target_title_hint": target_slide.title_hint if target_slide else None,
"required_points": target_slide.required_points if target_slide else [],
"judge_call_count": 0,
"used_previous_slide_context": previous_slide_extractions is not None,
"staleness": staleness,
"spec_hash": eval_spec.spec_hash,
"mode": mode,
},
)
__all__ = ["score_presentbench", "score_presentbench_slide"]
|