llm-data-eval-20261030 / scripts /gate_make_jsonl3.py
Raphael2099's picture
scripts
b2a67ba verified
Raw History Blame Contribute Delete
3.47 kB
# gate_make_jsonl3.py —— WW-6 gate stability probes (2026-10-30).
# Why: the 23:40 gate ran n=1 greedy per arm, so "3/3" was one draw, not stability.
# Four probes per arm, same six slips, same greedy decoding:
# P1 v2_repeat : byte-identical re-run of the 23:40 v2 prompt (determinism control)
# P2 v3_names : same 今译 task + explicit "proper nouns and numbers verbatim" constraint
# P3 v4_6463 : slips 63+64 only, one continuous translation (tests cross-slip carry in isolation)
# P4 v5_qa : targeted questions (subject of 日不食, origin of 三, meaning of 赏谷)
# 用法: python gate_make_jsonl3.py OUT.jsonl
import json, sys
SLIPS = [
("1", "吴王夫差起师伐越遂克越邦越王句践失邦移登会稽"),
("2", "之山使大夫种行成于吴曰寡君乏无使使非材种不敢当君天王"),
("3", "敢私告于下执事孤不天上帝降威懠于越邦不在前后当"),
("63", "邦人救火进者莫复王乃惧鼓而复之死者千人王大戚哭泣若主三"),
("64", "日不食乃训为之赏谷既成王亲以授其后既发王焉饮食王焉始命绝"),
("79", "宗庙凡吴之土地民人越君是尽既有之孤余何有面目以视于天下越君其事也"),
]
body = "".join("简%s:%s\n" % (n, t) for n, t in SLIPS)
body_63_64 = "".join("简%s:%s\n" % (n, t) for n, t in SLIPS if n in ("63", "64"))
P1 = (
"以下是出土简牍《吴王》篇六支简的校订释文(无句读,简与简按顺序连读)。\n"
"请按简号逐简给出今译,每简一行,只写译文,不要释文、不要说明。要求:紧贴简文,"
"只译简上有的词,不增补简上没有的信息;跨简连读的地方要按文意连起来译。\n\n"
+ body
)
P2 = (
"以下是出土简牍《吴王》篇六支简的校订释文(简与简按顺序连读)。请逐简给出今译,每简一行。\n"
"硬性要求:1) 人名、国名、地名必须与简文用字完全一致,简上写「越王」就写「越王」,"
"不得改成别的国名;2) 简上的数目字必须出现在译文对应处,不得省略;"
"3) 只译简上有的词,不增补简上没有的信息。\n\n"
+ body
)
P3 = (
"以下是出土简牍《吴王》篇相邻两支简的校订释文(无句读,两简连读)。\n"
"请把这两支简连成一段今译,只写译文。译完后另起一行写「注:」,"
"用一句话说明「赏谷」在这里指什么。\n\n"
+ body_63_64
)
P4 = (
"以下是出土简牍《吴王》篇相邻两支简的校订释文(无句读,两简连读):\n\n"
+ body_63_64 +
"\n请只回答三个问题,每问一行:\n"
"问1:简64「日不食」的主语是谁?\n"
"问2:简64 的「三」是从哪里来的?指什么?\n"
"问3:「赏谷」是什么意思?谁赏给谁?\n"
)
PROBES = [("P1_v2_repeat", P1, 900), ("P2_v3_names", P2, 900),
("P3_v4_6463", P3, 700), ("P4_v5_qa", P4, 500)]
rows = []
for pid, prompt, mn in PROBES:
rows.append({"id": pid, "max_new": mn,
"messages": [{"role": "user", "content": prompt}],
"slips": [{"slip_no": n, "text": t} for n, t in SLIPS]})
with open(sys.argv[1], "w", encoding="utf-8") as f:
for r in rows:
f.write(json.dumps(r, ensure_ascii=False) + "\n")
print("WROTE", sys.argv[1], "rows", len(rows), "chars", [len(p) for _, p, _ in PROBES])