Spaces:
Sleeping
Sleeping
File size: 21,114 Bytes
740e5d3 e33ee5a e6fc7db 01e58bc e6fc7db 01e58bc e6fc7db 01e58bc e6fc7db e33ee5a 740e5d3 e33ee5a 740e5d3 e6fc7db 01e58bc e6fc7db 740e5d3 e6fc7db 11240ae e6fc7db 740e5d3 e33ee5a e6fc7db 01e58bc e6fc7db 740e5d3 e33ee5a |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 |
import json
import io
from typing import List, Dict, Tuple
import pandas as pd
import numpy as np
import altair as alt
import streamlit as st
from pathlib import PurePosixPath
st.set_page_config(page_title="Twinkle Eval Analyzer", page_icon=":star2:", layout="wide")
st.title("✨ Twinkle Eval Analyzer (.json / .jsonl)")
# ----------------- Helpers -----------------
def _decode_bytes_to_text(b: bytes) -> str:
for enc in ("utf-8", "utf-16", "utf-16le", "utf-16be", "big5", "cp950"):
try:
return b.decode(enc)
except Exception:
continue
return b.decode("utf-8", errors="ignore")
def read_twinkle_doc(file) -> Dict:
raw = file.read()
if isinstance(raw, bytes):
text = _decode_bytes_to_text(raw)
else:
text = raw
text = text.strip()
try:
obj = json.loads(text)
except Exception:
for line in text.splitlines():
line = line.strip().rstrip(",")
if not line:
continue
try:
obj = json.loads(line)
break
except Exception:
continue
if not isinstance(obj, dict):
raise ValueError("檔案不是有效的 Twinkle Eval JSON 物件。")
if "timestamp" not in obj or "config" not in obj or "dataset_results" not in obj:
raise ValueError("缺少必要欄位")
return obj
def extract_records(doc: Dict) -> Tuple[pd.DataFrame, Dict[str, float]]:
model = doc.get("config", {}).get("model", {}).get("name", "<unknown>")
timestamp = doc.get("timestamp", "<no-ts>")
source_label = f"{model} @ {timestamp}"
rows = []
avg_map = {}
for ds_path, ds_payload in doc.get("dataset_results", {}).items():
ds_name = ds_path.split("datasets/")[-1].strip("/") if ds_path.startswith("datasets/") else ds_path
avg_meta = ds_payload.get("average_accuracy") if isinstance(ds_payload, dict) else None
results = ds_payload.get("results", []) if isinstance(ds_payload, dict) else []
for item in results:
if not isinstance(item, dict):
continue
file_path = item.get("file")
acc_mean = item.get("accuracy_mean")
if file_path is None or acc_mean is None:
continue
fname = PurePosixPath(file_path).name
category = fname.rsplit(".", 1)[0]
rows.append({
"dataset": ds_name,
"category": category,
"file": fname,
"accuracy_mean": float(acc_mean),
"source_label": source_label
})
if avg_meta is None and results:
vals = [float(it.get("accuracy_mean", np.nan)) for it in results if "accuracy_mean" in it]
if vals:
avg_meta = float(np.mean(vals))
if avg_meta is not None:
avg_map[ds_name] = avg_meta
return pd.DataFrame(rows), avg_map
def load_all(files) -> Tuple[pd.DataFrame, Dict[str, Dict[str, float]]]:
frames = []
meta = {}
for f in files or []:
try:
doc = read_twinkle_doc(f)
except Exception as e:
st.error(f"❌ 無法讀取 {getattr(f, 'name', '檔案')}:{e}")
continue
df, avg_map = extract_records(doc)
if not df.empty:
frames.append(df)
src = df["source_label"].iloc[0]
meta[src] = avg_map
if not frames:
return pd.DataFrame(columns=["dataset", "category", "file", "accuracy_mean", "source_label"]), {}
return pd.concat(frames, ignore_index=True), meta
# ----------------- Sidebar -----------------
with st.sidebar:
files = st.file_uploader("選擇 Twinkle Eval 檔案", type=["json", "jsonl"], accept_multiple_files=True)
df_all, meta_all = load_all(files)
normalize_0_100 = st.checkbox("以 0–100 顯示", value=False)
page_size = st.selectbox("每張圖顯示幾個類別", [10, 20, 30, 50, 100], index=1)
sort_mode = st.selectbox("排序方式(原始成績)", ["依整體平均由高到低", "依整體平均由低到高", "依字母排序"])
# === 分數閾值篩選 ===
st.markdown("---")
st.subheader("📏 分數篩選(原始成績)")
enable_threshold = st.checkbox("啟用分數閾值篩選", value=False)
if enable_threshold:
threshold_mode = st.radio("篩選模式", ["顯示 ≥ 閾值", "顯示 ≤ 閾值"])
# 篩選依據選擇
filter_basis = st.radio(
"篩選依據",
["任一模型符合", "特定模型符合"],
help="任一模型符合:只要有任一個模型在該類別符合條件就顯示\n特定模型符合:只篩選特定模型的分數"
)
specific_model = None
if filter_basis == "特定模型符合":
all_models = sorted(df_all["source_label"].unique().tolist()) if not df_all.empty else []
if all_models:
specific_model = st.selectbox("選擇模型", options=all_models)
# 根據顯示模式決定預設值與範圍
if normalize_0_100:
threshold_value = st.slider("閾值", min_value=0.0, max_value=100.0, value=50.0, step=1.0)
else:
threshold_value = st.slider("閾值", min_value=0.0, max_value=1.0, value=0.5, step=0.01)
st.caption(f"{'≥' if threshold_mode == '顯示 ≥ 閾值' else '≤'} {threshold_value}")
else:
threshold_mode = None
threshold_value = None
filter_basis = None
specific_model = None
# === Baseline Δ 圖表的控制 ===
st.markdown("---")
st.subheader("差距分析設定(Baseline Δ)")
options = ["|Δ| 由大到小", "Δ 由大到小(提升最多)", "Δ 由小到大(下降最多)", "依類別名稱"]
default = "Δ 由大到小(提升最多)"
delta_sort_mode = st.selectbox("差距排序方式(per-category)", options, index=options.index(default), key="delta_sort_mode")
abs_threshold = st.number_input("只顯示 |Δ| ≥ 門檻(可選)", min_value=0.0, value=0.0, step=0.1)
st.caption("Δ = Candidate 分數 − Baseline 分數;建議以 0–100 模式計算更直觀。")
if df_all.empty:
st.info("請上傳 Twinkle Eval 檔案")
st.stop()
# ----------------- 原始成績-----------------
all_datasets = sorted(df_all["dataset"].unique().tolist())
selected_dataset = st.selectbox("選擇資料集", options=all_datasets)
work = df_all[df_all["dataset"] == selected_dataset].copy()
metric_plot = "accuracy_mean" + (" (x100)" if normalize_0_100 else "")
work[metric_plot] = work["accuracy_mean"] * (100.0 if normalize_0_100 else 1.0)
# === 套用閾值篩選 ===
if enable_threshold and threshold_value is not None:
if filter_basis == "任一模型符合":
# 篩選方式:只要有任一模型在該類別符合條件
if threshold_mode == "顯示 ≥ 閾值":
# 保留那些「至少有一個模型 ≥ 閾值」的類別
valid_cats = work[work[metric_plot] >= threshold_value]["category"].unique().tolist()
else: # "顯示 ≤ 閾值"
# 保留那些「至少有一個模型 ≤ 閾值」的類別
valid_cats = work[work[metric_plot] <= threshold_value]["category"].unique().tolist()
work = work[work["category"].isin(valid_cats)]
# 計算平均用於排序
order_df = work.groupby("category")[metric_plot].mean().reset_index()
else: # "特定模型符合"
# 只篩選特定模型的資料,然後依該模型判斷
work_specific = work[work["source_label"] == specific_model]
if threshold_mode == "顯示 ≥ 閾值":
valid_cats = work_specific[work_specific[metric_plot] >= threshold_value]["category"].unique().tolist()
else: # "顯示 ≤ 閾值"
valid_cats = work_specific[work_specific[metric_plot] <= threshold_value]["category"].unique().tolist()
# 保留符合條件的類別,但顯示所有模型
work = work[work["category"].isin(valid_cats)]
order_df = work.groupby("category")[metric_plot].mean().reset_index()
else:
order_df = work.groupby("category")[metric_plot].mean().reset_index()
if sort_mode == "依整體平均由高到低":
order_df = order_df.sort_values(metric_plot, ascending=False)
elif sort_mode == "依整體平均由低到高":
order_df = order_df.sort_values(metric_plot, ascending=True)
else:
order_df = order_df.sort_values("category", ascending=True)
cat_order = order_df["category"].tolist()
# 如果篩選後沒有類別,顯示提示
if not cat_order:
if enable_threshold and threshold_mode is not None and threshold_value is not None:
st.warning(
f"⚠️ 沒有類別符合篩選條件({threshold_mode}: {threshold_value}),請調整閾值或關閉篩選。"
)
else:
st.warning("⚠️ 沒有可顯示的類別,請檢查資料集內容或調整篩選條件。")
st.stop()
work["category"] = pd.Categorical(work["category"], categories=cat_order, ordered=True)
n = len(cat_order)
pages = int(np.ceil(n / page_size))
st.markdown("## 📈 原始成績(各模型 × 類別)")
if enable_threshold and threshold_mode is not None:
filter_info = f"任一模型符合" if filter_basis == "任一模型符合" else f"模型 {specific_model} 符合"
st.info(f"🔍 已啟用篩選({filter_info}):顯示平均分數 {threshold_mode.replace('顯示 ', '')} {threshold_value} 的類別(共 {n} 個)")
for p in range(pages):
start, end = p * page_size, min((p + 1) * page_size, n)
subset_cats = cat_order[start:end]
sub = work[work["category"].isin(subset_cats)]
st.subheader(f"📊 {selected_dataset}|類別 {start+1}-{end} / {n}")
base = alt.Chart(sub).encode(
x=alt.X("category:N", sort=subset_cats),
y=alt.Y(f"{metric_plot}:Q"),
color=alt.Color("source_label:N"),
tooltip=["source_label", "file", alt.Tooltip(metric_plot, format=".3f")]
)
bars = base.mark_bar().encode(xOffset="source_label")
st.altair_chart(bars.properties(height=420), use_container_width=True)
pivot = sub.pivot_table(index="category", columns="source_label", values=metric_plot)
st.dataframe(pivot, use_container_width=True)
st.download_button(
label=f"下載此頁 CSV ({start+1}-{end})",
data=pivot.reset_index().to_csv(index=False).encode("utf-8"),
file_name=f"twinkle_{selected_dataset}_{start+1}_{end}.csv",
mime="text/csv"
)
# ----------------- 差距(Baseline Δ)分析 -----------------
st.markdown("---")
st.markdown("## ⚖️ 差距分析:Baseline vs. Candidates(Δ = Candidate − Baseline)")
# 使用與上方相同的資料集
dataset_for_delta = selected_dataset
df_delta_scope = df_all[df_all["dataset"] == dataset_for_delta].copy()
if df_delta_scope.empty:
st.warning(f"在資料集 **{dataset_for_delta}** 找不到資料,請確認上傳的 JSON 含此資料集名稱。")
try:
st.stop()
except Exception:
raise SystemExit
# 統一與上方尺度(建議用 0–100 再做差)
score_col = "score_0100"
df_delta_scope[score_col] = df_delta_scope["accuracy_mean"] * (100.0 if normalize_0_100 else 1.0)
# 手動指定 Baseline 與 Candidates
all_sources_in_scope = sorted(df_delta_scope["source_label"].unique().tolist())
col1, col2 = st.columns([1, 2])
with col1:
baseline = st.selectbox("選擇基準模型(Baseline)", options=all_sources_in_scope)
with col2:
default_candidates = [s for s in all_sources_in_scope if s != baseline]
candidates = st.multiselect("選擇要比較的候選模型(Candidates)", options=all_sources_in_scope, default=default_candidates)
if not candidates:
st.info("請至少選擇一個 Candidate。")
try:
st.stop()
except Exception:
raise SystemExit
# 建立寬表(index=category;已固定 dataset_for_delta)
wide = df_delta_scope.pivot_table(index="category", columns="source_label", values=score_col, aggfunc="mean")
# 只比較 baseline 與 candidates 的交集列
valid_candidates = [c for c in candidates if c in wide.columns]
if baseline not in wide.columns:
st.error("Baseline 在此資料集沒有任何分數可比。請換一個 Baseline 或資料集。")
try:
st.stop()
except Exception:
raise SystemExit
if not valid_candidates:
st.error("選取的 Candidates 在此資料集沒有任何分數可比。請換一組 Candidates 或資料集。")
try:
st.stop()
except Exception:
raise SystemExit
# 計算 Δ 長表(保留 baseline/candidate 原始分數)
delta_rows = []
for c in valid_candidates:
pair = wide[[baseline, c]].dropna() # 僅兩者皆有分數的類別
if pair.empty:
continue
for cat, row in pair.iterrows():
b = float(row[baseline])
s = float(row[c])
delta = s - b
if abs(delta) < abs_threshold: # 門檻過濾
continue
delta_rows.append({
"dataset": dataset_for_delta,
"category": cat,
"baseline": baseline,
"candidate": c,
"baseline_score": b,
"candidate_score": s,
"delta": delta
})
delta_df = pd.DataFrame(delta_rows)
if delta_df.empty:
st.warning("沒有符合條件的可比較類別(可能因缺漏或門檻過高)。")
try:
st.stop()
except Exception:
raise SystemExit
# 差距排序
if delta_sort_mode == "|Δ| 由大到小":
delta_df = delta_df.sort_values("delta", key=lambda s: s.abs(), ascending=False)
elif delta_sort_mode == "Δ 由大到小(提升最多)":
delta_df = delta_df.sort_values("delta", ascending=False)
elif delta_sort_mode == "Δ 由小到大(下降最多)":
delta_df = delta_df.sort_values("delta", ascending=True)
else:
delta_df = delta_df.sort_values("category", ascending=True)
# 圖表(Δ 不分頁,一次顯示全部類別)
tab1, tab2 = st.tabs(["📊 差距排行(per-category)", "📜 模型總結(per-candidate)"])
with tab1:
sub = delta_df.copy()
# === 先在 Pandas 內算出每個 candidate 的排序名次 ===
if delta_sort_mode == "Δ 由大到小(提升最多)":
sub["rank_in_candidate"] = sub.groupby("candidate")["delta"].rank(ascending=False, method="first")
table_sort = lambda df: df.sort_values(["candidate", "rank_in_candidate"], ascending=[True, True])
y_sort = alt.SortField("rank_in_candidate", order="ascending")
resolve_y = "independent"
elif delta_sort_mode == "Δ 由小到大(下降最多)":
sub["rank_in_candidate"] = sub.groupby("candidate")["delta"].rank(ascending=True, method="first")
table_sort = lambda df: df.sort_values(["candidate", "rank_in_candidate"], ascending=[True, True])
y_sort = alt.SortField("rank_in_candidate", order="ascending")
resolve_y = "independent"
elif delta_sort_mode == "|Δ| 由大到小":
sub["abs_delta"] = sub["delta"].abs()
sub["rank_in_candidate"] = sub.groupby("candidate")["abs_delta"].rank(ascending=False, method="first")
table_sort = lambda df: df.sort_values(["candidate", "rank_in_candidate"], ascending=[True, True])
y_sort = alt.SortField("rank_in_candidate", order="ascending")
resolve_y = "independent"
else: # 依類別名稱(字母序),共用排序
# 不用 rank,直接字母序
table_sort = lambda df: df.sort_values(["category", "candidate"], ascending=[True, True])
y_sort = alt.SortField("category", order="ascending")
resolve_y = "shared"
st.subheader(f"🔎 {dataset_for_delta}|Δ 排行(全部 {sub['category'].nunique()} 類別)")
chart_height = 25 * max(1, sub["category"].nunique())
base = alt.Chart(sub).encode(
y=alt.Y("category:N", sort=y_sort, title="Category"),
x=alt.X("delta:Q", title="Δ = Candidate − Baseline"),
color=alt.Color("candidate:N", title="Candidate"),
tooltip=[
alt.Tooltip("category:N", title="Category"),
alt.Tooltip("candidate:N", title="Candidate"),
alt.Tooltip("baseline:N", title="Baseline"),
alt.Tooltip("baseline_score:Q", title="Baseline 分數", format=".3f"),
alt.Tooltip("candidate_score:Q", title="Candidate 分數", format=".3f"),
alt.Tooltip("delta:Q", title="Δ", format=".3f"),
],
)
chart = (
base.mark_bar()
.encode(row=alt.Row("candidate:N", header=alt.Header(title=None)))
.properties(height=chart_height)
.resolve_scale(y=resolve_y) # 各 candidate 分面各自排序或共用
)
st.altair_chart(chart, use_container_width=True)
# 表格:依 rank_in_candidate 排序,與圖一致
table = table_sort(sub)[["category", "candidate", "baseline_score", "candidate_score", "delta"]]
st.dataframe(table, use_container_width=True)
st.download_button(
label="下載 Δ 排行 CSV(全部類別)",
data=table.to_csv(index=False).encode("utf-8"),
file_name=f"delta_{dataset_for_delta}_ALL.csv",
mime="text/csv",
)
with tab2:
# per-candidate 總結:mean/median Δ、win/lose/tie、覆蓋率、Top/Bottom-N
summaries = []
top_k = st.number_input("Top/Bottom-N(顯示每個 Candidate 的最大/最小差距分類)", min_value=1, value=10, step=1)
for c in valid_candidates:
pair = wide[[baseline, c]].dropna()
if pair.empty:
continue
deltas = pair[c] - pair[baseline]
m = float(np.mean(deltas))
med = float(np.median(deltas))
win = int((deltas > 0).sum())
lose = int((deltas < 0).sum())
tie = int((deltas == 0).sum())
coverage = f"{len(deltas)}/{wide.shape[0]}" # 有共同分數的類別數 / 全部類別數
# 取 Top/Bottom-N 類別(按 Δ)
top_rows = (pair.assign(delta=deltas)
.sort_values("delta", ascending=False)
.head(top_k)
.reset_index()[["category", baseline, c, "delta"]])
bottom_rows = (pair.assign(delta=deltas)
.sort_values("delta", ascending=True)
.head(top_k)
.reset_index()[["category", baseline, c, "delta"]])
summaries.append({
"candidate": c,
"mean_delta": m,
"median_delta": med,
"win": win,
"lose": lose,
"tie": tie,
"coverage": coverage,
"top_list": top_rows,
"bottom_list": bottom_rows
})
if not summaries:
st.warning("沒有可用的 per-candidate 總結(可能都沒有交集)。")
else:
# 概覽表
overview = pd.DataFrame([{
"Candidate": s["candidate"],
"Mean Δ": s["mean_delta"],
"Median Δ": s["median_delta"],
"Win": s["win"],
"Lose": s["lose"],
"Tie": s["tie"],
"Coverage (交集/總類別)": s["coverage"],
} for s in summaries]).sort_values("Mean Δ", ascending=False)
st.markdown("### 總覽(與 Baseline 成對比較)")
st.dataframe(overview, use_container_width=True)
st.download_button(
label="下載 per-candidate 總覽 CSV",
data=overview.to_csv(index=False).encode("utf-8"),
file_name=f"delta_overview_{dataset_for_delta}.csv",
mime="text/csv"
)
# 逐 Candidate 顯示 Top/Bottom-N 清單(可收合)
st.markdown("### 各 Candidate 的差距清單(Top/Bottom-N)")
for s in summaries:
with st.expander(f"🔸 {s['candidate']}"):
st.write("**Top-N(提升最多)**")
top_tbl = s["top_list"].rename(columns={baseline: "baseline_score", s["candidate"]: "candidate_score"})
st.dataframe(top_tbl, use_container_width=True)
st.download_button(
label=f"下載 {s['candidate']} Top-N",
data=top_tbl.to_csv(index=False).encode("utf-8"),
file_name=f"delta_top_{dataset_for_delta}_{s['candidate']}.csv",
mime="text/csv"
)
st.write("**Bottom-N(下降最多)**")
bottom_tbl = s["bottom_list"].rename(columns={baseline: "baseline_score", s["candidate"]: "candidate_score"})
st.dataframe(bottom_tbl, use_container_width=True)
st.download_button(
label=f"下載 {s['candidate']} Bottom-N",
data=bottom_tbl.to_csv(index=False).encode("utf-8"),
file_name=f"delta_bottom_{dataset_for_delta}_{s['candidate']}.csv",
mime="text/csv"
)
|