support arbitrary K for estimated_position (top1/top5/top100)
Browse files- add_estimated_position.py +63 -40
add_estimated_position.py
CHANGED
|
@@ -1,16 +1,25 @@
|
|
| 1 |
#!/usr/bin/env python3
|
| 2 |
-
"""把
|
| 3 |
|
| 4 |
-
|
| 5 |
-
- 从 golden_top100.jsonl 取它的 100 个最近邻 ruler 的 rank
|
| 6 |
-
- 求均值,round 成整数
|
| 7 |
-
- 作为新列 estimated_position 写回 csv
|
| 8 |
-
|
| 9 |
-
输出:golden_set_with_estimated.csv(不覆盖原文件)
|
| 10 |
|
| 11 |
用法:
|
| 12 |
-
|
| 13 |
-
python3 add_estimated_position.py --
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
"""
|
| 15 |
import argparse
|
| 16 |
import json
|
|
@@ -25,78 +34,92 @@ DEFAULTS = dict(
|
|
| 25 |
jsonl = "golden_top100.jsonl",
|
| 26 |
output = "golden_set_with_estimated.csv",
|
| 27 |
id_col = "task_id",
|
| 28 |
-
new_col = "estimated_position",
|
| 29 |
)
|
| 30 |
|
| 31 |
|
| 32 |
def parse_args():
|
| 33 |
p = argparse.ArgumentParser()
|
| 34 |
-
p.add_argument("--csv",
|
| 35 |
-
p.add_argument("--jsonl",
|
| 36 |
-
p.add_argument("--output",
|
| 37 |
-
p.add_argument("--id-col",
|
| 38 |
-
p.add_argument("--
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 39 |
return p.parse_args()
|
| 40 |
|
| 41 |
|
| 42 |
def main():
|
| 43 |
args = parse_args()
|
|
|
|
| 44 |
|
| 45 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
print(f"[1] read jsonl: {args.jsonl}")
|
| 47 |
estimated = {}
|
|
|
|
| 48 |
with open(args.jsonl, "r", encoding="utf-8") as f:
|
| 49 |
for line in f:
|
| 50 |
line = line.strip()
|
| 51 |
if not line:
|
| 52 |
continue
|
| 53 |
r = json.loads(line)
|
| 54 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 55 |
estimated[str(r["task_id"])] = int(round(ranks.mean()))
|
| 56 |
-
print(f" -> {len(estimated)} cases estimated")
|
|
|
|
|
|
|
| 57 |
|
| 58 |
# 2) 读 csv
|
| 59 |
print(f"[2] read csv: {args.csv}")
|
| 60 |
df = pd.read_csv(args.csv, keep_default_na=False)
|
| 61 |
-
print(f" -> {len(df)} rows")
|
| 62 |
|
| 63 |
if args.id_col not in df.columns:
|
| 64 |
raise ValueError(f"'{args.id_col}' not in csv. cols: {list(df.columns)}")
|
| 65 |
|
| 66 |
-
# 3) 给每条 case 加
|
| 67 |
-
print(f"[3] add column: {
|
| 68 |
-
df[
|
| 69 |
|
| 70 |
-
n_total
|
| 71 |
-
n_matched = df[
|
| 72 |
-
n_missing = n_total - n_matched
|
| 73 |
print(f" matched: {n_matched}/{n_total}")
|
| 74 |
-
if
|
| 75 |
-
|
| 76 |
-
else:
|
| 77 |
-
# 全部匹配上 → 转成 Int64(pandas 的可空整数类型,不会因为没 NaN 变 float)
|
| 78 |
-
df[args.new_col] = df[args.new_col].astype("Int64")
|
| 79 |
|
| 80 |
# 4) 写
|
| 81 |
print(f"[4] write: {args.output}")
|
| 82 |
df.to_csv(args.output, index=False)
|
| 83 |
-
print(f" done. csv 现在有 {len(df.columns)} 列(
|
| 84 |
|
| 85 |
-
# 5) 预览
|
| 86 |
-
print(f"\n[preview] 前 5 行:")
|
| 87 |
-
cols = [args.id_col, "label",
|
| 88 |
cols = [c for c in cols if c in df.columns]
|
| 89 |
print(df[cols].head(5).to_string(index=False))
|
| 90 |
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
print(f"
|
| 94 |
-
print(f" min={int(s.min())} max={int(s.max())} mean={s.mean():.1f} median={int(s.median())}")
|
| 95 |
if "label" in df.columns:
|
| 96 |
print("\n按 label 分组:")
|
| 97 |
for lab, sub in df.groupby("label"):
|
| 98 |
-
ss = pd.to_numeric(sub[
|
| 99 |
-
print(f" label={lab!r:>4} n={len(sub):>4} mean={ss.mean():.
|
| 100 |
|
| 101 |
|
| 102 |
if __name__ == "__main__":
|
|
|
|
| 1 |
#!/usr/bin/env python3
|
| 2 |
+
"""把 embedding top-K match 估出来的整数 position 写回 golden_set.csv。
|
| 3 |
|
| 4 |
+
支持任意 K(默认 100),列名自动带 K 后缀,方便同时跑多种 K 比较。
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
|
| 6 |
用法:
|
| 7 |
+
# top-100 邻居均值
|
| 8 |
+
python3 add_estimated_position.py --k 100
|
| 9 |
+
|
| 10 |
+
# top-5 邻居均值
|
| 11 |
+
python3 add_estimated_position.py --k 5
|
| 12 |
+
|
| 13 |
+
# top-1(最近邻 rank)
|
| 14 |
+
python3 add_estimated_position.py --k 1
|
| 15 |
+
|
| 16 |
+
# 一次跑多种 K(每种生成一列)
|
| 17 |
+
python3 add_estimated_position.py --k 1 --k-suffix top1
|
| 18 |
+
python3 add_estimated_position.py --csv golden_set_with_estimated.csv --k 5 --k-suffix top5
|
| 19 |
+
python3 add_estimated_position.py --csv golden_set_with_estimated.csv --k 100 --k-suffix top100
|
| 20 |
+
|
| 21 |
+
# 自定义列名
|
| 22 |
+
python3 add_estimated_position.py --k 5 --new-col estimated_position_top5
|
| 23 |
"""
|
| 24 |
import argparse
|
| 25 |
import json
|
|
|
|
| 34 |
jsonl = "golden_top100.jsonl",
|
| 35 |
output = "golden_set_with_estimated.csv",
|
| 36 |
id_col = "task_id",
|
|
|
|
| 37 |
)
|
| 38 |
|
| 39 |
|
| 40 |
def parse_args():
|
| 41 |
p = argparse.ArgumentParser()
|
| 42 |
+
p.add_argument("--csv", default=DEFAULTS["csv"])
|
| 43 |
+
p.add_argument("--jsonl", default=DEFAULTS["jsonl"])
|
| 44 |
+
p.add_argument("--output", default=DEFAULTS["output"])
|
| 45 |
+
p.add_argument("--id-col", default=DEFAULTS["id_col"])
|
| 46 |
+
p.add_argument("--k", type=int, default=100,
|
| 47 |
+
help="用前 K 个邻居算均值(K=1 即 top1 rank)")
|
| 48 |
+
p.add_argument("--new-col", default=None,
|
| 49 |
+
help="新列名;不传则自动用 estimated_position_top{K}")
|
| 50 |
+
p.add_argument("--k-suffix", default=None,
|
| 51 |
+
help="(已弃用,等价于 --new-col estimated_position_{suffix})")
|
| 52 |
return p.parse_args()
|
| 53 |
|
| 54 |
|
| 55 |
def main():
|
| 56 |
args = parse_args()
|
| 57 |
+
K = args.k
|
| 58 |
|
| 59 |
+
if args.new_col:
|
| 60 |
+
col_name = args.new_col
|
| 61 |
+
elif args.k_suffix:
|
| 62 |
+
col_name = f"estimated_position_{args.k_suffix}"
|
| 63 |
+
else:
|
| 64 |
+
col_name = f"estimated_position_top{K}"
|
| 65 |
+
|
| 66 |
+
# 1) 读 jsonl,每条样本算 top-K rank 均值(K=1 时等价于最近邻 rank)
|
| 67 |
print(f"[1] read jsonl: {args.jsonl}")
|
| 68 |
estimated = {}
|
| 69 |
+
skipped = 0
|
| 70 |
with open(args.jsonl, "r", encoding="utf-8") as f:
|
| 71 |
for line in f:
|
| 72 |
line = line.strip()
|
| 73 |
if not line:
|
| 74 |
continue
|
| 75 |
r = json.loads(line)
|
| 76 |
+
topk = r.get("topk") or []
|
| 77 |
+
if len(topk) < K:
|
| 78 |
+
skipped += 1
|
| 79 |
+
continue
|
| 80 |
+
ranks = np.array([t["rank"] for t in topk[:K]], dtype=float)
|
| 81 |
estimated[str(r["task_id"])] = int(round(ranks.mean()))
|
| 82 |
+
print(f" -> {len(estimated)} cases estimated (K={K})")
|
| 83 |
+
if skipped:
|
| 84 |
+
print(f" ⚠️ {skipped} 条 jsonl 邻居数 < K,跳过")
|
| 85 |
|
| 86 |
# 2) 读 csv
|
| 87 |
print(f"[2] read csv: {args.csv}")
|
| 88 |
df = pd.read_csv(args.csv, keep_default_na=False)
|
| 89 |
+
print(f" -> {len(df)} rows, {len(df.columns)} cols")
|
| 90 |
|
| 91 |
if args.id_col not in df.columns:
|
| 92 |
raise ValueError(f"'{args.id_col}' not in csv. cols: {list(df.columns)}")
|
| 93 |
|
| 94 |
+
# 3) 给每条 case 加新列
|
| 95 |
+
print(f"[3] add column: {col_name}")
|
| 96 |
+
df[col_name] = df[args.id_col].astype(str).map(estimated)
|
| 97 |
|
| 98 |
+
n_total = len(df)
|
| 99 |
+
n_matched = df[col_name].notna().sum()
|
|
|
|
| 100 |
print(f" matched: {n_matched}/{n_total}")
|
| 101 |
+
if n_matched == n_total:
|
| 102 |
+
df[col_name] = df[col_name].astype("Int64")
|
|
|
|
|
|
|
|
|
|
| 103 |
|
| 104 |
# 4) 写
|
| 105 |
print(f"[4] write: {args.output}")
|
| 106 |
df.to_csv(args.output, index=False)
|
| 107 |
+
print(f" done. csv 现在有 {len(df.columns)} 列(含新列:{col_name})")
|
| 108 |
|
| 109 |
+
# 5) 预览 + stats
|
| 110 |
+
print(f"\n[preview] 前 5 行(task_id / label / {col_name}):")
|
| 111 |
+
cols = [args.id_col, "label", col_name]
|
| 112 |
cols = [c for c in cols if c in df.columns]
|
| 113 |
print(df[cols].head(5).to_string(index=False))
|
| 114 |
|
| 115 |
+
s = pd.to_numeric(df[col_name], errors="coerce")
|
| 116 |
+
print(f"\n[stats] {col_name}:")
|
| 117 |
+
print(f" min={int(s.min())} max={int(s.max())} mean={s.mean():.2f} median={int(s.median())} std={s.std():.2f}")
|
|
|
|
| 118 |
if "label" in df.columns:
|
| 119 |
print("\n按 label 分组:")
|
| 120 |
for lab, sub in df.groupby("label"):
|
| 121 |
+
ss = pd.to_numeric(sub[col_name], errors="coerce")
|
| 122 |
+
print(f" label={lab!r:>4} n={len(sub):>4} mean={ss.mean():.2f} median={int(ss.median())} std={ss.std():.2f}")
|
| 123 |
|
| 124 |
|
| 125 |
if __name__ == "__main__":
|