Spaces:
Running
Running
小形克宏 commited on
Commit ·
5ba08e6
1
Parent(s): f614727
Fix: replace Dataframe with Markdown to avoid Gradio bug
Browse files
app.py
CHANGED
|
@@ -1,9 +1,6 @@
|
|
| 1 |
"""
|
| 2 |
StructEval-T Analyzer
|
| 3 |
松尾研LLM講義2025 メインコンペ用 推論結果分析ツール
|
| 4 |
-
|
| 5 |
-
inference.json と public_150.json をアップロードして、
|
| 6 |
-
フォーマット別のパース成功率やエラーパターンを分析します。
|
| 7 |
"""
|
| 8 |
|
| 9 |
import json
|
|
@@ -18,32 +15,25 @@ import gradio as gr
|
|
| 18 |
import pandas as pd
|
| 19 |
|
| 20 |
# ---------------------------------------------------------------------------
|
| 21 |
-
# 1. Syntax Validators
|
| 22 |
# ---------------------------------------------------------------------------
|
| 23 |
|
| 24 |
-
def validate_json(text
|
| 25 |
-
"""JSON構文を検証"""
|
| 26 |
try:
|
| 27 |
json.loads(text)
|
| 28 |
return True, ""
|
| 29 |
except json.JSONDecodeError as e:
|
| 30 |
-
return False, f"JSONDecodeError: {e.msg} (line {e.lineno}
|
| 31 |
-
|
| 32 |
|
| 33 |
-
def validate_yaml(text
|
| 34 |
-
"""YAML構文を検証"""
|
| 35 |
try:
|
| 36 |
import yaml
|
| 37 |
yaml.safe_load(text)
|
| 38 |
return True, ""
|
| 39 |
-
except yaml.YAMLError as e:
|
| 40 |
-
return False, f"YAMLError: {e}"
|
| 41 |
except Exception as e:
|
| 42 |
-
return False, f"
|
| 43 |
-
|
| 44 |
|
| 45 |
-
def validate_toml(text
|
| 46 |
-
"""TOML構文を検証"""
|
| 47 |
try:
|
| 48 |
import tomllib
|
| 49 |
tomllib.loads(text)
|
|
@@ -51,37 +41,29 @@ def validate_toml(text: str) -> tuple[bool, str]:
|
|
| 51 |
except Exception as e:
|
| 52 |
return False, f"TOMLError: {e}"
|
| 53 |
|
| 54 |
-
|
| 55 |
-
def validate_xml(text: str) -> tuple[bool, str]:
|
| 56 |
-
"""XML構文を検証"""
|
| 57 |
try:
|
| 58 |
import xml.etree.ElementTree as ET
|
| 59 |
ET.fromstring(text)
|
| 60 |
return True, ""
|
| 61 |
-
except ET.ParseError as e:
|
| 62 |
-
return False, f"XMLParseError: {e}"
|
| 63 |
except Exception as e:
|
| 64 |
-
return False, f"
|
| 65 |
-
|
| 66 |
|
| 67 |
-
def validate_csv(text
|
| 68 |
-
"""CSV構文を検証"""
|
| 69 |
try:
|
| 70 |
reader = csv.reader(io.StringIO(text))
|
| 71 |
rows = list(reader)
|
| 72 |
if len(rows) == 0:
|
| 73 |
return False, "Empty CSV"
|
| 74 |
if len(rows) == 1:
|
| 75 |
-
return False, "
|
| 76 |
-
# 列数の一貫性チェック
|
| 77 |
col_counts = [len(row) for row in rows]
|
| 78 |
if len(set(col_counts)) > 1:
|
| 79 |
-
return False, f"Inconsistent
|
| 80 |
return True, ""
|
| 81 |
except Exception as e:
|
| 82 |
return False, f"CSVError: {e}"
|
| 83 |
|
| 84 |
-
|
| 85 |
VALIDATORS = {
|
| 86 |
"JSON": validate_json,
|
| 87 |
"YAML": validate_yaml,
|
|
@@ -91,112 +73,63 @@ VALIDATORS = {
|
|
| 91 |
}
|
| 92 |
|
| 93 |
# ---------------------------------------------------------------------------
|
| 94 |
-
# 2. Error Pattern Classifier
|
| 95 |
# ---------------------------------------------------------------------------
|
| 96 |
|
| 97 |
-
def classify_error_patterns(generation
|
| 98 |
-
"""出力テキストのエラーパターンを分類"""
|
| 99 |
patterns = []
|
| 100 |
-
|
| 101 |
-
# マークダウンブロックの混入
|
| 102 |
if re.search(r"```\w*", generation):
|
| 103 |
patterns.append("markdown_block")
|
| 104 |
-
|
| 105 |
-
# 自然言語の混入(先頭部分)
|
| 106 |
first_line = generation.strip().split("\n")[0] if generation.strip() else ""
|
| 107 |
-
nl_indicators = [
|
| 108 |
-
"here is", "here's", "below is", "the following",
|
| 109 |
-
"sure", "certainly", "of course", "i'll",
|
| 110 |
-
"let me", "note:", "output:",
|
| 111 |
-
]
|
| 112 |
if any(ind in first_line.lower() for ind in nl_indicators):
|
| 113 |
patterns.append("natural_language_prefix")
|
| 114 |
-
|
| 115 |
-
# 末尾の自然言語混入
|
| 116 |
last_lines = generation.strip().split("\n")[-3:] if generation.strip() else []
|
| 117 |
last_text = " ".join(last_lines).lower()
|
| 118 |
-
|
| 119 |
-
if any(ind in last_text for ind in nl_suffix):
|
| 120 |
patterns.append("natural_language_suffix")
|
| 121 |
-
|
| 122 |
-
# 途切れ(トランケーション)の検出
|
| 123 |
-
stripped = generation.rstrip()
|
| 124 |
if output_type == "JSON":
|
| 125 |
-
|
| 126 |
-
close_count = generation.count("}") + generation.count("]")
|
| 127 |
-
if open_count > close_count:
|
| 128 |
patterns.append("truncation")
|
| 129 |
elif output_type == "XML":
|
| 130 |
open_tags = len(re.findall(r"<[^/!?][^>]*>", generation))
|
| 131 |
close_tags = len(re.findall(r"</[^>]+>", generation))
|
| 132 |
if open_tags > close_tags + 1:
|
| 133 |
patterns.append("truncation")
|
| 134 |
-
elif output_type in ("YAML", "TOML", "CSV"):
|
| 135 |
-
if stripped and stripped[-1] == "\\":
|
| 136 |
-
patterns.append("truncation")
|
| 137 |
-
|
| 138 |
-
# 空出力
|
| 139 |
if not generation.strip():
|
| 140 |
patterns.append("empty_output")
|
| 141 |
-
|
| 142 |
-
# 別フォーマットの出力(JSONを要求されたのにXMLが出てくる等)
|
| 143 |
-
format_indicators = {
|
| 144 |
-
"JSON": (r"^\s*[\{\[]", None),
|
| 145 |
-
"XML": (r"^\s*<", None),
|
| 146 |
-
"YAML": (None, None),
|
| 147 |
-
"TOML": (r"^\s*\[", None),
|
| 148 |
-
"CSV": (None, None),
|
| 149 |
-
}
|
| 150 |
-
if output_type == "JSON" and re.match(r"^\s*<", generation.strip()):
|
| 151 |
-
patterns.append("wrong_format")
|
| 152 |
-
elif output_type == "XML" and re.match(r"^\s*[\{\[]", generation.strip()):
|
| 153 |
-
patterns.append("wrong_format")
|
| 154 |
-
|
| 155 |
-
# CoT思考過程の混入
|
| 156 |
-
if re.search(r"<think>|</think>|<reasoning>|</reasoning>", generation):
|
| 157 |
patterns.append("cot_leakage")
|
| 158 |
-
|
|
|
|
| 159 |
return patterns if patterns else ["unknown"]
|
| 160 |
|
| 161 |
-
|
| 162 |
# ---------------------------------------------------------------------------
|
| 163 |
-
# 3. Core Analysis
|
| 164 |
# ---------------------------------------------------------------------------
|
| 165 |
|
| 166 |
-
def load_public_150(file_path
|
| 167 |
-
"""public_150.json を読み込み、task_id → 情報 の辞書を返す"""
|
| 168 |
with open(file_path, "r", encoding="utf-8") as f:
|
| 169 |
data = json.load(f)
|
| 170 |
return {item["task_id"]: item for item in data}
|
| 171 |
|
| 172 |
-
|
| 173 |
-
def analyze_single_inference(
|
| 174 |
-
inference_data: list[dict],
|
| 175 |
-
task_info: dict,
|
| 176 |
-
) -> pd.DataFrame:
|
| 177 |
-
"""1つのinference.jsonを分析してDataFrameを返す"""
|
| 178 |
results = []
|
| 179 |
for item in inference_data:
|
| 180 |
task_id = item.get("task_id", "")
|
| 181 |
generation = item.get("generation", "")
|
| 182 |
-
|
| 183 |
info = task_info.get(task_id, {})
|
| 184 |
output_type = info.get("output_type", "UNKNOWN")
|
| 185 |
task_name = info.get("task_name", "UNKNOWN")
|
| 186 |
-
|
| 187 |
-
# 構文検証
|
| 188 |
validator = VALIDATORS.get(output_type)
|
| 189 |
if validator:
|
| 190 |
is_valid, error_msg = validator(generation)
|
| 191 |
else:
|
| 192 |
is_valid, error_msg = False, f"Unknown format: {output_type}"
|
| 193 |
-
|
| 194 |
-
# エラーパターン分類
|
| 195 |
if not is_valid:
|
| 196 |
error_patterns = classify_error_patterns(generation, output_type)
|
| 197 |
else:
|
| 198 |
error_patterns = []
|
| 199 |
-
|
| 200 |
results.append({
|
| 201 |
"task_id": task_id,
|
| 202 |
"task_name": task_name,
|
|
@@ -207,91 +140,48 @@ def analyze_single_inference(
|
|
| 207 |
"generation_length": len(generation),
|
| 208 |
"generation_preview": generation[:200],
|
| 209 |
})
|
| 210 |
-
|
| 211 |
return pd.DataFrame(results)
|
| 212 |
|
| 213 |
-
|
| 214 |
-
def compute_summary(df: pd.DataFrame) -> dict:
|
| 215 |
-
"""分析結果のサマリーを計算"""
|
| 216 |
total = len(df)
|
| 217 |
-
valid = df["is_valid"].sum()
|
| 218 |
-
|
| 219 |
summary = {
|
| 220 |
"total_tasks": total,
|
| 221 |
-
"parse_success":
|
| 222 |
-
"parse_fail":
|
| 223 |
"parse_rate": f"{valid / total * 100:.1f}%" if total > 0 else "N/A",
|
| 224 |
}
|
| 225 |
-
|
| 226 |
-
# フォーマット別
|
| 227 |
format_stats = {}
|
| 228 |
for fmt in ["JSON", "YAML", "TOML", "XML", "CSV"]:
|
| 229 |
fmt_df = df[df["output_type"] == fmt]
|
| 230 |
fmt_total = len(fmt_df)
|
| 231 |
-
fmt_valid = fmt_df["is_valid"].sum()
|
| 232 |
format_stats[fmt] = {
|
| 233 |
"total": fmt_total,
|
| 234 |
-
"success":
|
| 235 |
-
"fail":
|
| 236 |
"rate": f"{fmt_valid / fmt_total * 100:.1f}%" if fmt_total > 0 else "N/A",
|
| 237 |
}
|
| 238 |
summary["by_format"] = format_stats
|
| 239 |
-
|
| 240 |
-
# エラーパターン集計
|
| 241 |
all_patterns = []
|
| 242 |
for patterns_str in df[df["is_valid"] == False]["error_patterns"]:
|
| 243 |
if patterns_str:
|
| 244 |
all_patterns.extend(patterns_str.split(","))
|
| 245 |
summary["error_pattern_counts"] = dict(Counter(all_patterns).most_common())
|
| 246 |
-
|
| 247 |
return summary
|
| 248 |
|
| 249 |
-
|
| 250 |
# ---------------------------------------------------------------------------
|
| 251 |
-
# 4.
|
| 252 |
-
# ---------------------------------------------------------------------------
|
| 253 |
-
|
| 254 |
-
def compare_experiments(
|
| 255 |
-
all_results: dict[str, pd.DataFrame],
|
| 256 |
-
) -> pd.DataFrame:
|
| 257 |
-
"""複数実験の結果を比較するDataFrameを返す"""
|
| 258 |
-
rows = []
|
| 259 |
-
for name, df in all_results.items():
|
| 260 |
-
total = len(df)
|
| 261 |
-
valid = df["is_valid"].sum()
|
| 262 |
-
row = {
|
| 263 |
-
"experiment": name,
|
| 264 |
-
"total": total,
|
| 265 |
-
"parse_success": int(valid),
|
| 266 |
-
"parse_rate": f"{valid / total * 100:.1f}%" if total > 0 else "N/A",
|
| 267 |
-
}
|
| 268 |
-
for fmt in ["JSON", "YAML", "TOML", "XML", "CSV"]:
|
| 269 |
-
fmt_df = df[df["output_type"] == fmt]
|
| 270 |
-
fmt_total = len(fmt_df)
|
| 271 |
-
fmt_valid = fmt_df["is_valid"].sum()
|
| 272 |
-
row[f"{fmt}_rate"] = (
|
| 273 |
-
f"{fmt_valid / fmt_total * 100:.1f}%"
|
| 274 |
-
if fmt_total > 0
|
| 275 |
-
else "N/A"
|
| 276 |
-
)
|
| 277 |
-
rows.append(row)
|
| 278 |
-
return pd.DataFrame(rows)
|
| 279 |
-
|
| 280 |
-
|
| 281 |
-
# ---------------------------------------------------------------------------
|
| 282 |
-
# 5. Gradio Interface
|
| 283 |
# ---------------------------------------------------------------------------
|
| 284 |
|
| 285 |
def process_files(public_150_file, inference_files):
|
| 286 |
-
"""メイン処理:ファイルを受け取って分析結果を返す"""
|
| 287 |
if public_150_file is None:
|
| 288 |
-
return "❌ public_150.json をアップロードしてください",
|
| 289 |
|
| 290 |
if not inference_files:
|
| 291 |
-
return "❌ inference.json を1つ以上アップロードしてください",
|
| 292 |
|
| 293 |
try:
|
| 294 |
-
# Gradio 5ではfilepathモードで文字列パスが渡される
|
| 295 |
pub_path = public_150_file if isinstance(public_150_file, str) else public_150_file.name
|
| 296 |
task_info = load_public_150(pub_path)
|
| 297 |
|
|
@@ -303,13 +193,12 @@ def process_files(public_150_file, inference_files):
|
|
| 303 |
filename = Path(inf_path).stem
|
| 304 |
with open(inf_path, "r", encoding="utf-8") as f:
|
| 305 |
inference_data = json.load(f)
|
| 306 |
-
|
| 307 |
df = analyze_single_inference(inference_data, task_info)
|
| 308 |
summary = compute_summary(df)
|
| 309 |
all_results[filename] = df
|
| 310 |
all_summaries[filename] = summary
|
| 311 |
|
| 312 |
-
# ---
|
| 313 |
summary_text = "## 📊 分析結果サマリー\n\n"
|
| 314 |
for name, s in all_summaries.items():
|
| 315 |
summary_text += f"### {name}\n"
|
|
@@ -323,127 +212,114 @@ def process_files(public_150_file, inference_files):
|
|
| 323 |
summary_text += f" - {pattern}: {count}件\n"
|
| 324 |
summary_text += "\n"
|
| 325 |
|
| 326 |
-
# ---
|
| 327 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 328 |
|
| 329 |
-
# ---
|
| 330 |
first_name = list(all_results.keys())[0]
|
| 331 |
first_df = all_results[first_name]
|
| 332 |
-
error_df = first_df[first_df["is_valid"] == False]
|
| 333 |
-
["task_id", "task_name", "output_type", "error_msg", "error_patterns", "generation_preview"]
|
| 334 |
-
]
|
| 335 |
|
| 336 |
-
|
| 337 |
-
|
| 338 |
-
|
| 339 |
-
|
| 340 |
-
|
| 341 |
-
|
| 342 |
-
|
| 343 |
-
|
| 344 |
-
row[fmt] = round(fmt_valid / fmt_total * 100, 1) if fmt_total > 0 else 0
|
| 345 |
-
format_comparison_rows.append(row)
|
| 346 |
-
format_df = pd.DataFrame(format_comparison_rows)
|
| 347 |
|
| 348 |
-
return summary_text,
|
| 349 |
|
| 350 |
except Exception as e:
|
| 351 |
error_trace = traceback.format_exc()
|
| 352 |
-
return f"❌ エラー
|
| 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 |
-
type="filepath",
|
| 384 |
-
)
|
| 385 |
-
inference_files = gr.File(
|
| 386 |
-
label="inference.json(複数可)",
|
| 387 |
-
file_types=[".json"],
|
| 388 |
-
file_count="multiple",
|
| 389 |
-
type="filepath",
|
| 390 |
-
)
|
| 391 |
-
|
| 392 |
-
analyze_btn = gr.Button("🔬 分析開始", variant="primary", size="lg")
|
| 393 |
-
|
| 394 |
-
with gr.Tabs():
|
| 395 |
-
with gr.Tab("📊 サマリー"):
|
| 396 |
-
summary_output = gr.Markdown()
|
| 397 |
-
|
| 398 |
-
with gr.Tab("📈 実験比較"):
|
| 399 |
-
comparison_table = gr.Dataframe(
|
| 400 |
-
label="実験間のパース成功率比較",
|
| 401 |
-
interactive=False,
|
| 402 |
-
)
|
| 403 |
-
|
| 404 |
-
with gr.Tab("❌ エラー詳細"):
|
| 405 |
-
gr.Markdown("*最初にアップロードされたファイルのエラー一覧を表示*")
|
| 406 |
-
error_table = gr.Dataframe(
|
| 407 |
-
label="パース失敗タスク一覧",
|
| 408 |
-
interactive=False,
|
| 409 |
-
wrap=True,
|
| 410 |
-
)
|
| 411 |
-
|
| 412 |
-
with gr.Tab("📉 フォーマット別"):
|
| 413 |
-
format_table = gr.Dataframe(
|
| 414 |
-
label="フォーマット別パース成功率(%)",
|
| 415 |
-
interactive=False,
|
| 416 |
-
)
|
| 417 |
-
|
| 418 |
-
analyze_btn.click(
|
| 419 |
-
fn=process_files,
|
| 420 |
-
inputs=[public_file, inference_files],
|
| 421 |
-
outputs=[summary_output, comparison_table, error_table, format_table],
|
| 422 |
)
|
| 423 |
-
|
| 424 |
-
|
| 425 |
-
""
|
| 426 |
-
|
| 427 |
-
|
| 428 |
-
運営側の採点基準である `raw_output_metric`(特定キーの存在チェック等)は
|
| 429 |
-
`public_150.json` から削除されているため、完全なスコア再現はできません。
|
| 430 |
-
|
| 431 |
-
**エラーパターンの凡例:**
|
| 432 |
-
- `markdown_block`: マークダウンコードブロック(\\`\\`\\`json 等)の混入
|
| 433 |
-
- `natural_language_prefix`: 先頭に自然言語("Here is..."等)が混入
|
| 434 |
-
- `natural_language_suffix`: 末尾に自然言語("Note:"等)が混入
|
| 435 |
-
- `truncation`: 出力の途切れ(閉じ括弧・タグの欠落)
|
| 436 |
-
- `empty_output`: 空の出力
|
| 437 |
-
- `wrong_format`: 要求と異なるフォーマットの出力
|
| 438 |
-
- `cot_leakage`: 思考過程(\\<think\\>等)の混入
|
| 439 |
-
- `unknown`: 上記に該当しない構文エラー
|
| 440 |
-
"""
|
| 441 |
)
|
| 442 |
|
| 443 |
-
|
| 444 |
-
|
| 445 |
-
|
| 446 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 447 |
|
| 448 |
if __name__ == "__main__":
|
| 449 |
-
demo.launch()
|
|
|
|
| 1 |
"""
|
| 2 |
StructEval-T Analyzer
|
| 3 |
松尾研LLM講義2025 メインコンペ用 推論結果分析ツール
|
|
|
|
|
|
|
|
|
|
| 4 |
"""
|
| 5 |
|
| 6 |
import json
|
|
|
|
| 15 |
import pandas as pd
|
| 16 |
|
| 17 |
# ---------------------------------------------------------------------------
|
| 18 |
+
# 1. Syntax Validators
|
| 19 |
# ---------------------------------------------------------------------------
|
| 20 |
|
| 21 |
+
def validate_json(text):
|
|
|
|
| 22 |
try:
|
| 23 |
json.loads(text)
|
| 24 |
return True, ""
|
| 25 |
except json.JSONDecodeError as e:
|
| 26 |
+
return False, f"JSONDecodeError: {e.msg} (line {e.lineno})"
|
|
|
|
| 27 |
|
| 28 |
+
def validate_yaml(text):
|
|
|
|
| 29 |
try:
|
| 30 |
import yaml
|
| 31 |
yaml.safe_load(text)
|
| 32 |
return True, ""
|
|
|
|
|
|
|
| 33 |
except Exception as e:
|
| 34 |
+
return False, f"YAMLError: {e}"
|
|
|
|
| 35 |
|
| 36 |
+
def validate_toml(text):
|
|
|
|
| 37 |
try:
|
| 38 |
import tomllib
|
| 39 |
tomllib.loads(text)
|
|
|
|
| 41 |
except Exception as e:
|
| 42 |
return False, f"TOMLError: {e}"
|
| 43 |
|
| 44 |
+
def validate_xml(text):
|
|
|
|
|
|
|
| 45 |
try:
|
| 46 |
import xml.etree.ElementTree as ET
|
| 47 |
ET.fromstring(text)
|
| 48 |
return True, ""
|
|
|
|
|
|
|
| 49 |
except Exception as e:
|
| 50 |
+
return False, f"XMLError: {e}"
|
|
|
|
| 51 |
|
| 52 |
+
def validate_csv(text):
|
|
|
|
| 53 |
try:
|
| 54 |
reader = csv.reader(io.StringIO(text))
|
| 55 |
rows = list(reader)
|
| 56 |
if len(rows) == 0:
|
| 57 |
return False, "Empty CSV"
|
| 58 |
if len(rows) == 1:
|
| 59 |
+
return False, "Only header"
|
|
|
|
| 60 |
col_counts = [len(row) for row in rows]
|
| 61 |
if len(set(col_counts)) > 1:
|
| 62 |
+
return False, f"Inconsistent cols: {set(col_counts)}"
|
| 63 |
return True, ""
|
| 64 |
except Exception as e:
|
| 65 |
return False, f"CSVError: {e}"
|
| 66 |
|
|
|
|
| 67 |
VALIDATORS = {
|
| 68 |
"JSON": validate_json,
|
| 69 |
"YAML": validate_yaml,
|
|
|
|
| 73 |
}
|
| 74 |
|
| 75 |
# ---------------------------------------------------------------------------
|
| 76 |
+
# 2. Error Pattern Classifier
|
| 77 |
# ---------------------------------------------------------------------------
|
| 78 |
|
| 79 |
+
def classify_error_patterns(generation, output_type):
|
|
|
|
| 80 |
patterns = []
|
|
|
|
|
|
|
| 81 |
if re.search(r"```\w*", generation):
|
| 82 |
patterns.append("markdown_block")
|
|
|
|
|
|
|
| 83 |
first_line = generation.strip().split("\n")[0] if generation.strip() else ""
|
| 84 |
+
nl_indicators = ["here is", "here's", "below is", "sure", "certainly", "let me"]
|
|
|
|
|
|
|
|
|
|
|
|
|
| 85 |
if any(ind in first_line.lower() for ind in nl_indicators):
|
| 86 |
patterns.append("natural_language_prefix")
|
|
|
|
|
|
|
| 87 |
last_lines = generation.strip().split("\n")[-3:] if generation.strip() else []
|
| 88 |
last_text = " ".join(last_lines).lower()
|
| 89 |
+
if any(ind in last_text for ind in ["note:", "explanation:", "this ", "the above"]):
|
|
|
|
| 90 |
patterns.append("natural_language_suffix")
|
|
|
|
|
|
|
|
|
|
| 91 |
if output_type == "JSON":
|
| 92 |
+
if generation.count("{") + generation.count("[") > generation.count("}") + generation.count("]"):
|
|
|
|
|
|
|
| 93 |
patterns.append("truncation")
|
| 94 |
elif output_type == "XML":
|
| 95 |
open_tags = len(re.findall(r"<[^/!?][^>]*>", generation))
|
| 96 |
close_tags = len(re.findall(r"</[^>]+>", generation))
|
| 97 |
if open_tags > close_tags + 1:
|
| 98 |
patterns.append("truncation")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 99 |
if not generation.strip():
|
| 100 |
patterns.append("empty_output")
|
| 101 |
+
if re.search(r"<think>|</think>", generation):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 102 |
patterns.append("cot_leakage")
|
| 103 |
+
if re.search(r"<tool_call>", generation):
|
| 104 |
+
patterns.append("tool_call_leakage")
|
| 105 |
return patterns if patterns else ["unknown"]
|
| 106 |
|
|
|
|
| 107 |
# ---------------------------------------------------------------------------
|
| 108 |
+
# 3. Core Analysis
|
| 109 |
# ---------------------------------------------------------------------------
|
| 110 |
|
| 111 |
+
def load_public_150(file_path):
|
|
|
|
| 112 |
with open(file_path, "r", encoding="utf-8") as f:
|
| 113 |
data = json.load(f)
|
| 114 |
return {item["task_id"]: item for item in data}
|
| 115 |
|
| 116 |
+
def analyze_single_inference(inference_data, task_info):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 117 |
results = []
|
| 118 |
for item in inference_data:
|
| 119 |
task_id = item.get("task_id", "")
|
| 120 |
generation = item.get("generation", "")
|
|
|
|
| 121 |
info = task_info.get(task_id, {})
|
| 122 |
output_type = info.get("output_type", "UNKNOWN")
|
| 123 |
task_name = info.get("task_name", "UNKNOWN")
|
|
|
|
|
|
|
| 124 |
validator = VALIDATORS.get(output_type)
|
| 125 |
if validator:
|
| 126 |
is_valid, error_msg = validator(generation)
|
| 127 |
else:
|
| 128 |
is_valid, error_msg = False, f"Unknown format: {output_type}"
|
|
|
|
|
|
|
| 129 |
if not is_valid:
|
| 130 |
error_patterns = classify_error_patterns(generation, output_type)
|
| 131 |
else:
|
| 132 |
error_patterns = []
|
|
|
|
| 133 |
results.append({
|
| 134 |
"task_id": task_id,
|
| 135 |
"task_name": task_name,
|
|
|
|
| 140 |
"generation_length": len(generation),
|
| 141 |
"generation_preview": generation[:200],
|
| 142 |
})
|
|
|
|
| 143 |
return pd.DataFrame(results)
|
| 144 |
|
| 145 |
+
def compute_summary(df):
|
|
|
|
|
|
|
| 146 |
total = len(df)
|
| 147 |
+
valid = int(df["is_valid"].sum())
|
|
|
|
| 148 |
summary = {
|
| 149 |
"total_tasks": total,
|
| 150 |
+
"parse_success": valid,
|
| 151 |
+
"parse_fail": total - valid,
|
| 152 |
"parse_rate": f"{valid / total * 100:.1f}%" if total > 0 else "N/A",
|
| 153 |
}
|
|
|
|
|
|
|
| 154 |
format_stats = {}
|
| 155 |
for fmt in ["JSON", "YAML", "TOML", "XML", "CSV"]:
|
| 156 |
fmt_df = df[df["output_type"] == fmt]
|
| 157 |
fmt_total = len(fmt_df)
|
| 158 |
+
fmt_valid = int(fmt_df["is_valid"].sum())
|
| 159 |
format_stats[fmt] = {
|
| 160 |
"total": fmt_total,
|
| 161 |
+
"success": fmt_valid,
|
| 162 |
+
"fail": fmt_total - fmt_valid,
|
| 163 |
"rate": f"{fmt_valid / fmt_total * 100:.1f}%" if fmt_total > 0 else "N/A",
|
| 164 |
}
|
| 165 |
summary["by_format"] = format_stats
|
|
|
|
|
|
|
| 166 |
all_patterns = []
|
| 167 |
for patterns_str in df[df["is_valid"] == False]["error_patterns"]:
|
| 168 |
if patterns_str:
|
| 169 |
all_patterns.extend(patterns_str.split(","))
|
| 170 |
summary["error_pattern_counts"] = dict(Counter(all_patterns).most_common())
|
|
|
|
| 171 |
return summary
|
| 172 |
|
|
|
|
| 173 |
# ---------------------------------------------------------------------------
|
| 174 |
+
# 4. Main Processing
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 175 |
# ---------------------------------------------------------------------------
|
| 176 |
|
| 177 |
def process_files(public_150_file, inference_files):
|
|
|
|
| 178 |
if public_150_file is None:
|
| 179 |
+
return "❌ public_150.json をアップロードしてください", "", ""
|
| 180 |
|
| 181 |
if not inference_files:
|
| 182 |
+
return "❌ inference.json を1つ以上アップロードしてください", "", ""
|
| 183 |
|
| 184 |
try:
|
|
|
|
| 185 |
pub_path = public_150_file if isinstance(public_150_file, str) else public_150_file.name
|
| 186 |
task_info = load_public_150(pub_path)
|
| 187 |
|
|
|
|
| 193 |
filename = Path(inf_path).stem
|
| 194 |
with open(inf_path, "r", encoding="utf-8") as f:
|
| 195 |
inference_data = json.load(f)
|
|
|
|
| 196 |
df = analyze_single_inference(inference_data, task_info)
|
| 197 |
summary = compute_summary(df)
|
| 198 |
all_results[filename] = df
|
| 199 |
all_summaries[filename] = summary
|
| 200 |
|
| 201 |
+
# --- Output 1: Summary ---
|
| 202 |
summary_text = "## 📊 分析結果サマリー\n\n"
|
| 203 |
for name, s in all_summaries.items():
|
| 204 |
summary_text += f"### {name}\n"
|
|
|
|
| 212 |
summary_text += f" - {pattern}: {count}件\n"
|
| 213 |
summary_text += "\n"
|
| 214 |
|
| 215 |
+
# --- Output 2: Comparison table as markdown ---
|
| 216 |
+
comp_lines = ["## 📈 実験比較\n"]
|
| 217 |
+
comp_lines.append("| experiment | total | pass | rate | JSON | YAML | TOML | XML | CSV |")
|
| 218 |
+
comp_lines.append("|---|---|---|---|---|---|---|---|---|")
|
| 219 |
+
for name, df in all_results.items():
|
| 220 |
+
total = len(df)
|
| 221 |
+
valid = int(df["is_valid"].sum())
|
| 222 |
+
rate = f"{valid/total*100:.1f}%" if total > 0 else "N/A"
|
| 223 |
+
fmt_rates = {}
|
| 224 |
+
for fmt in ["JSON", "YAML", "TOML", "XML", "CSV"]:
|
| 225 |
+
fmt_df = df[df["output_type"] == fmt]
|
| 226 |
+
ft = len(fmt_df)
|
| 227 |
+
fv = int(fmt_df["is_valid"].sum())
|
| 228 |
+
fmt_rates[fmt] = f"{fv/ft*100:.1f}%" if ft > 0 else "N/A"
|
| 229 |
+
comp_lines.append(f"| {name} | {total} | {valid} | {rate} | {fmt_rates['JSON']} | {fmt_rates['YAML']} | {fmt_rates['TOML']} | {fmt_rates['XML']} | {fmt_rates['CSV']} |")
|
| 230 |
+
comparison_md = "\n".join(comp_lines)
|
| 231 |
|
| 232 |
+
# --- Output 3: Error details as markdown ---
|
| 233 |
first_name = list(all_results.keys())[0]
|
| 234 |
first_df = all_results[first_name]
|
| 235 |
+
error_df = first_df[first_df["is_valid"] == False]
|
|
|
|
|
|
|
| 236 |
|
| 237 |
+
error_lines = [f"## ❌ エラー詳細 ({first_name})\n"]
|
| 238 |
+
error_lines.append(f"パース失敗: {len(error_df)}件\n")
|
| 239 |
+
error_lines.append("| task_name | output_type | error_patterns | error_msg |")
|
| 240 |
+
error_lines.append("|---|---|---|---|")
|
| 241 |
+
for _, row in error_df.iterrows():
|
| 242 |
+
err_msg_short = str(row['error_msg'])[:60]
|
| 243 |
+
error_lines.append(f"| {row['task_name']} | {row['output_type']} | {row['error_patterns']} | {err_msg_short} |")
|
| 244 |
+
error_md = "\n".join(error_lines)
|
|
|
|
|
|
|
|
|
|
| 245 |
|
| 246 |
+
return summary_text, comparison_md, error_md
|
| 247 |
|
| 248 |
except Exception as e:
|
| 249 |
error_trace = traceback.format_exc()
|
| 250 |
+
return f"❌ エラー:\n```\n{error_trace}\n```", "", ""
|
|
|
|
| 251 |
|
| 252 |
# ---------------------------------------------------------------------------
|
| 253 |
+
# 5. Gradio App - using only Markdown outputs to avoid Dataframe bugs
|
| 254 |
# ---------------------------------------------------------------------------
|
| 255 |
|
| 256 |
+
with gr.Blocks(
|
| 257 |
+
title="StructEval-T Analyzer",
|
| 258 |
+
theme=gr.themes.Soft(),
|
| 259 |
+
) as demo:
|
| 260 |
+
gr.Markdown(
|
| 261 |
+
"""
|
| 262 |
+
# 🔍 StructEval-T Analyzer
|
| 263 |
+
### 松尾研LLM講義2025 メインコンペ用 推論結果分析ツール
|
| 264 |
+
|
| 265 |
+
`inference.json` と `public_150.json` をアップロードすることで、
|
| 266 |
+
モデル出力の構文的正確性(パース可能性)やエラーパターンを分析できます。
|
| 267 |
+
|
| 268 |
+
**使い方:**
|
| 269 |
+
1. `public_150.json` をアップロード
|
| 270 |
+
2. 1つ以上の `inference.json` をアップロード(複数ファイル対応・実験比較可能)
|
| 271 |
+
3. 「分析開始」ボタンをクリック
|
| 272 |
+
"""
|
| 273 |
+
)
|
| 274 |
+
|
| 275 |
+
with gr.Row():
|
| 276 |
+
public_file = gr.File(
|
| 277 |
+
label="public_150.json",
|
| 278 |
+
file_types=[".json"],
|
| 279 |
+
type="filepath",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 280 |
)
|
| 281 |
+
inference_files = gr.File(
|
| 282 |
+
label="inference.json(��数可)",
|
| 283 |
+
file_types=[".json"],
|
| 284 |
+
file_count="multiple",
|
| 285 |
+
type="filepath",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 286 |
)
|
| 287 |
|
| 288 |
+
analyze_btn = gr.Button("🔬 分析開始", variant="primary", size="lg")
|
| 289 |
+
|
| 290 |
+
with gr.Tabs():
|
| 291 |
+
with gr.Tab("📊 サマリー"):
|
| 292 |
+
summary_output = gr.Markdown()
|
| 293 |
+
with gr.Tab("📈 実験比較"):
|
| 294 |
+
comparison_output = gr.Markdown()
|
| 295 |
+
with gr.Tab("❌ エラー詳細"):
|
| 296 |
+
error_output = gr.Markdown()
|
| 297 |
+
|
| 298 |
+
analyze_btn.click(
|
| 299 |
+
fn=process_files,
|
| 300 |
+
inputs=[public_file, inference_files],
|
| 301 |
+
outputs=[summary_output, comparison_output, error_output],
|
| 302 |
+
)
|
| 303 |
+
|
| 304 |
+
gr.Markdown(
|
| 305 |
+
"""
|
| 306 |
+
---
|
| 307 |
+
**注意:** このツールは構文的な正確性(パース可能かどうか)のみを検証します。
|
| 308 |
+
運営側の採点基準である `raw_output_metric`(特定キーの存在チェック等)は
|
| 309 |
+
`public_150.json` から削除されているため、完全なスコア再現はできません。
|
| 310 |
+
|
| 311 |
+
**エラーパターンの凡例:**
|
| 312 |
+
- `markdown_block`: マークダウンコードブロック(\\`\\`\\`json 等)の混入
|
| 313 |
+
- `natural_language_prefix`: 先頭に自然言語("Here is..."等)が混入
|
| 314 |
+
- `natural_language_suffix`: 末尾に自然言語("Note:"等)が混入
|
| 315 |
+
- `truncation`: 出力の途切れ(閉じ括弧・タグの欠落)
|
| 316 |
+
- `empty_output`: 空の出力
|
| 317 |
+
- `wrong_format`: 要求と異なるフォーマットの出力
|
| 318 |
+
- `cot_leakage`: 思考過程(\\<think\\>等)の混入
|
| 319 |
+
- `tool_call_leakage`: ツールコール(\\<tool_call\\>等)の混入
|
| 320 |
+
- `unknown`: 上記に該当しない構文エラー
|
| 321 |
+
"""
|
| 322 |
+
)
|
| 323 |
|
| 324 |
if __name__ == "__main__":
|
| 325 |
+
demo.launch(ssr=False)
|