File size: 9,391 Bytes
df455fc ba2cf79 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 | from __future__ import annotations
import hashlib
import json
from ..reward_models import (
ChecklistItem,
PresentBenchEvalSpec,
TaskSpec,
to_serializable,
)
SPEC_VERSION = "1.0"
DEFAULT_PRESENTBENCH_SCORING_CONFIG = {
"dimension_weights": {
"fundamentals": 0.20,
"visual_layout": 0.20,
"completeness": 0.20,
"correctness": 0.20,
"fidelity": 0.20,
},
"aesthetic_weight": 0.15,
"hard_caps": {
"blank_title_only_ratio_threshold": 0.4,
},
"soft_penalties": {
"slide_count_violation": 0.02,
"overlap": 0.01,
"tiny_text": 0.01,
"redundancy": 0.03,
"wrong_slot_behavior": 0.02,
"staleness": 0.30,
},
}
def generate_checklist(task_spec: TaskSpec) -> list[ChecklistItem]:
checklist: list[ChecklistItem] = [
ChecklistItem(
item_id="fundamentals_slide_count",
dimension="fundamentals",
prompt_text="Is the slide count within the requested range?",
item_kind="slide_count_range",
),
ChecklistItem(
item_id="fundamentals_theme",
dimension="fundamentals",
prompt_text="Does the deck maintain a central theme aligned with the prompt?",
item_kind="theme_alignment",
relevant_sections=task_spec.required_sections,
),
ChecklistItem(
item_id="fundamentals_audience",
dimension="fundamentals",
prompt_text="Is the deck appropriate for the intended audience and tone?",
item_kind="audience_tone",
),
ChecklistItem(
item_id="visual_readability",
dimension="visual_layout",
prompt_text="Is the text readable without tiny fonts?",
item_kind="readable_text",
),
ChecklistItem(
item_id="visual_overlap",
dimension="visual_layout",
prompt_text="Are there no major overlaps or clipping risks?",
item_kind="no_major_overlap",
),
ChecklistItem(
item_id="visual_consistency",
dimension="visual_layout",
prompt_text="Is the visual design reasonably consistent across slides?",
item_kind="design_consistency",
),
]
for index, section in enumerate(task_spec.required_sections, start=1):
checklist.append(
ChecklistItem(
item_id=f"completeness_section_{index:02d}",
dimension="completeness",
prompt_text=f"Does the deck include a clear '{section}' section?",
item_kind="required_section",
relevant_sections=[section],
)
)
for index, point in enumerate(task_spec.required_points, start=1):
source_refs = [
fact["ref"]
for fact in task_spec.metadata.get("source_facts", [])
if point in fact["text"]
][:2]
checklist.append(
ChecklistItem(
item_id=f"completeness_point_{index:02d}",
dimension="completeness",
prompt_text=f"Does the deck cover this required point: {point}?",
item_kind="required_point",
source_refs=source_refs,
)
)
checklist.append(
ChecklistItem(
item_id=f"correctness_point_{index:02d}",
dimension="correctness",
prompt_text=f"Is this required point stated correctly: {point}?",
item_kind="correct_required_point",
source_refs=source_refs,
)
)
if task_spec.required_slides:
for slide in task_spec.required_slides:
checklist.append(
ChecklistItem(
item_id=f"fidelity_slide_{slide.slide_index:02d}",
dimension="fidelity",
prompt_text=f"Is all content on Slide {slide.slide_index} supported by the source pack?",
item_kind="slide_fidelity",
required_slide_scope=[slide.slide_index],
relevant_sections=[slide.slide_role],
)
)
else:
checklist.append(
ChecklistItem(
item_id="fidelity_deck",
dimension="fidelity",
prompt_text="Is all deck content supported by the source pack?",
item_kind="deck_fidelity",
)
)
return checklist
def generate_slide_checklists(task_spec: TaskSpec) -> dict[int, list[ChecklistItem]]:
slide_checklists: dict[int, list[ChecklistItem]] = {}
for slide in task_spec.required_slides or []:
items = [
ChecklistItem(
item_id=f"slide_{slide.slide_index:02d}_prompt_alignment",
dimension="prompt_alignment",
prompt_text=f"Does Slide {slide.slide_index} match the intended role '{slide.slide_role}'?",
item_kind="slide_role_match",
required_slide_scope=[slide.slide_index],
relevant_sections=[slide.slide_role],
),
ChecklistItem(
item_id=f"slide_{slide.slide_index:02d}_title_alignment",
dimension="prompt_alignment",
prompt_text=f"Is Slide {slide.slide_index} title aligned with '{slide.title_hint or slide.instructions}'?",
item_kind="slide_title_alignment",
required_slide_scope=[slide.slide_index],
relevant_sections=[slide.slide_role],
),
ChecklistItem(
item_id=f"slide_{slide.slide_index:02d}_fidelity",
dimension="local_fidelity",
prompt_text=f"Is all content on Slide {slide.slide_index} supported by the source pack?",
item_kind="slide_fidelity",
required_slide_scope=[slide.slide_index],
relevant_sections=[slide.slide_role],
),
ChecklistItem(
item_id=f"slide_{slide.slide_index:02d}_usability",
dimension="local_usability",
prompt_text=f"Is Slide {slide.slide_index} readable and free of major clutter?",
item_kind="slide_readability",
required_slide_scope=[slide.slide_index],
relevant_sections=[slide.slide_role],
),
]
for point_index, point in enumerate(slide.required_points, start=1):
items.append(
ChecklistItem(
item_id=f"slide_{slide.slide_index:02d}_point_{point_index:02d}",
dimension="local_completeness",
prompt_text=f"Does Slide {slide.slide_index} cover this point: {point}?",
item_kind="slide_required_point",
required_slide_scope=[slide.slide_index],
relevant_sections=[slide.slide_role],
)
)
for exact_index, exact_value in enumerate(slide.required_exact_values, start=1):
items.append(
ChecklistItem(
item_id=f"slide_{slide.slide_index:02d}_exact_{exact_index:02d}",
dimension="local_correctness",
prompt_text=f"Does Slide {slide.slide_index} include the exact value '{exact_value}' correctly?",
item_kind="slide_exact_value",
required_slide_scope=[slide.slide_index],
relevant_sections=[slide.slide_role],
)
)
if slide.required_shape_kinds:
items.append(
ChecklistItem(
item_id=f"slide_{slide.slide_index:02d}_visual_kind",
dimension="local_completeness",
prompt_text=f"Does Slide {slide.slide_index} include the expected supported visual forms?",
item_kind="slide_required_visual",
required_slide_scope=[slide.slide_index],
relevant_sections=[slide.slide_role],
)
)
slide_checklists[slide.slide_index] = items
return slide_checklists
def build_presentbench_eval_spec(
task_spec: TaskSpec, *, mode: str = "eval"
) -> PresentBenchEvalSpec:
checklist = generate_checklist(task_spec)
slide_checklists = generate_slide_checklists(task_spec)
payload = {
"task_spec": to_serializable(task_spec),
"checklist": to_serializable(checklist),
"slide_checklists": to_serializable(slide_checklists),
"scoring_config": DEFAULT_PRESENTBENCH_SCORING_CONFIG,
"spec_version": SPEC_VERSION,
}
spec_hash = hashlib.sha256(
json.dumps(payload, sort_keys=True, separators=(",", ":")).encode("utf-8")
).hexdigest()
return PresentBenchEvalSpec(
task_spec=task_spec,
checklist=checklist,
slide_checklists=slide_checklists,
scoring_config={**DEFAULT_PRESENTBENCH_SCORING_CONFIG, "mode": mode},
spec_version=SPEC_VERSION,
spec_hash=spec_hash,
)
__all__ = [
"DEFAULT_PRESENTBENCH_SCORING_CONFIG",
"SPEC_VERSION",
"build_presentbench_eval_spec",
"generate_checklist",
"generate_slide_checklists",
]
|