File size: 7,419 Bytes
79c6440 | 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 | # -*- coding: utf-8 -*-
"""Vibrato 情绪专项题库 — 类 Jev 契约(score + choice + noul 三题型) v0.1.1
Laya 的三种题型(qtype)在情绪域的完整特化:
score ×3 P/A/D × 9档李克特bin (契约在 pad_schema.py, 此处引用)
choice 8 族训练头 + OCC-22 距离排序(occ.py 解码层输出)
noul ×2 文本层才能判的是非信号 (宿主行为信号, 非情绪动力学)
v0.1.1 变更:
- noul 5→2: 保留 suppressed(言不由衷) + directed_at_me(指向性)——
这两题 PAD 推不出来(定义性不可导), 只能靠文本判读;
negative(P<0 弱导出 0.78) / needs_comfort / escalating(宿主有ΔPAD历史自己算更准)
移除——宿主或 LLM 自行判断更可靠
- route 字段改名 host_signal, 去掉 chordia 路由错误表述(Chordia 是 agent 情绪
动力学引擎, 不是行为路由器)
冻结纪律同 pad_schema: 题干/选项/顺序改动 = 契约破裂, battery_digest() 两侧比对。
teacher 标注: 一条消息一次调用, 18项+族+是非 合并一个 JSON(省 6 倍调用)。
"""
from __future__ import annotations
import hashlib
import json
from typing import Any, Dict, List, Optional
import pad_schema
# ── choice: 情绪族(8) ───────────────────────────────────────
# Laya EMOTION_BUCKETS 7 族 + 委屈(中文伴侣聊天特有, 英文体系缺位)
FAMILY_LABELS: List[Dict[str, str]] = [
{"id": "happy", "zh": "开心"},
{"id": "affectionate", "zh": "亲昵"},
{"id": "calm", "zh": "平静"},
{"id": "amused", "zh": "好笑"},
{"id": "sad", "zh": "低落"},
{"id": "anxious", "zh": "焦虑"},
{"id": "angry", "zh": "生气"},
{"id": "wronged", "zh": "委屈"},
]
FAMILY_IDS: List[str] = [f["id"] for f in FAMILY_LABELS]
N_FAMILY = len(FAMILY_IDS)
FAMILY_INSTRUCTIONS = (
"TARGET 消息的发送者当前的主导情绪是哪一族？只选一个。"
"开心=愉悦满足；亲昵=对对方的喜爱亲近；平静=事务性/无明显情绪；"
"好笑=被逗乐/玩梗；低落=难过沮丧疲惫；焦虑=担心紧张不安；"
"生气=愤怒不满恼火；委屈=觉得被误解/被亏待/有苦说不出。"
)
# ── noul: 文本层才能判的是非信号(2) ──────────────────────
# 这两题 PAD 推不出来(定义性不可导), 只能靠文本判读:
# suppressed: 表里错位(为真时表面PAD≠真实PAD, 一个PAD判不了"装")
# directed_at_me: 指代只在文本里("对你失望"vs"对老板失望" PAD 逐位相同)
# id, 中文问题, 判定说明(teacher 用), 宿主行为信号
NOUL_QUESTIONS: List[Dict[str, str]] = [
{"id": "directed_at_me", "q": "情绪是否指向对话中的我",
"def": "情绪的靶子是接收者(我), 而非第三方/自身处境/泛泛而谈",
"host_signal": "真→宿主可能需要道歉或担责"},
{"id": "suppressed", "q": "是否言不由衷/压抑",
"def": "表面措辞与真实情绪不符(敷衍/逞强/反话/故作轻松)则为真",
"host_signal": "真→宿主轻问深挖, 勿按字面回应"},
]
NOUL_IDS: List[str] = [q["id"] for q in NOUL_QUESTIONS]
N_NOUL = len(NOUL_IDS)
def battery_digest() -> str:
canonical = json.dumps(
{"pad": pad_schema.contract_digest(),
"family": {"ids": FAMILY_IDS, "instructions": FAMILY_INSTRUCTIONS},
"noul": [{"id": q["id"], "q": q["q"]} for q in NOUL_QUESTIONS]},
ensure_ascii=False, sort_keys=True, separators=(",", ":"))
return hashlib.sha256(canonical.encode("utf-8")).hexdigest()
def build_teacher_prompt(state: str, prev_pad: Optional[List[float]] = None) -> str:
"""情绪专项合并标注提示词: 一次调用出 18项 + 情绪族 + 5是非。"""
lines = [
"你是情绪标注专家。判断下面 TARGET 消息发送者的情绪状态, 只输出 JSON。",
"", state, "",
"回答三部分(合成一个 JSON 对象, 禁止解释文字):",
"1. I1-I18: PAD 情感量表 18 词项, 1.00-9.00 两位小数。",
" P维 I1快乐-I2高兴-I3满意-I4惬意-I5希望-I6放松(9为符合,5中立,1反面);",
" A维 I7兴奋-I8警觉-I9刺激-I10狂热-I11活跃-I12惊慌;",
" D维 I13支配-I14影响-I15领导-I16重要-I17自由-I18强力。",
f"2. family: 主导情绪族, 只能取: {', '.join(FAMILY_IDS)}。",
"3. noul: 两个是非判断(true/false):",
]
for q in NOUL_QUESTIONS:
lines.append(f" {q['id']}: {q['def']}")
lines += [
"",
'输出格式: {"I1":..,"I18":..,"analysis":"..","family":"..",',
' "noul":{"directed_at_me":true,"suppressed":false}}',
]
return "\n".join(lines)
def parse_teacher_json(raw: str) -> Dict[str, Any]:
"""teacher 输出 → {items: {I1..I18}, family: id, noul: {id: bool}}。校验+裁剪。
词项分越界(如 0.0 或归一化值混入)一律钳位到 [1,9] 而非拒收——0.0 多为
模型把量表当 0-9 用(即极度负向), 钳位保信息保 echo 链连续。
"""
text = raw.strip()
if text.startswith("```"):
text = text.strip("`").removeprefix("json").strip()
data = json.loads(text)
items = {}
for i in range(1, 19):
v = float(data[f"I{i}"])
items[f"I{i}"] = max(1.0, min(9.0, v))
family = data["family"]
if family not in FAMILY_IDS:
raise ValueError(f"未知情绪族: {family}")
noul_raw = data["noul"]
noul = {qid: bool(noul_raw[qid]) for qid in NOUL_IDS}
return {"items": items, "family": family, "noul": noul, "analysis": data.get("analysis", "")}
def validate_battery_gold(gold: Dict[str, Any]) -> List[str]:
problems: List[str] = list(pad_schema.validate_gold(gold))
if gold.get("family") not in FAMILY_IDS:
problems.append(f"family 非法: {gold.get('family')}")
for qid in NOUL_IDS:
if not isinstance(gold.get("noul", {}).get(qid), bool):
problems.append(f"noul.{qid} 缺失或非布尔")
return problems
if __name__ == "__main__":
# 题库自检
assert N_FAMILY == 8 and N_NOUL == 2 and len(set(FAMILY_IDS)) == N_FAMILY
assert set(NOUL_IDS) == {"directed_at_me", "suppressed"}
prompt = build_teacher_prompt("TARGET message to judge:\n[them]: 哼, 随便你吧")
for token in ["I18", "family", "wronged", "suppressed", "directed_at_me"]:
assert token in prompt, token
parsed = parse_teacher_json(json.dumps({
**{f"I{i}": 5.0 for i in range(1, 19)}, "family": "wronged",
"noul": {q: True for q in NOUL_IDS}}))
assert parsed["family"] == "wronged" and all(parsed["noul"].values())
assert pad_schema.bins_to_pad(pad_schema.items_to_bins(parsed["items"]))["pleasure"] == 0.0
good = {"pad_bins": pad_schema.items_to_bins({f"I{i}": 5.0 for i in range(1, 19)}),
"echo": [0.0] * 3, "family": "calm",
"noul": {q: False for q in NOUL_IDS}}
assert validate_battery_gold(good) == []
bad = {"pad_bins": good["pad_bins"], "echo": [0.0] * 3, "family": "happy", "noul": {}}
assert len(validate_battery_gold(bad)) == N_NOUL
print("battery_digest:", battery_digest())
print("题库: score×3(P/A/D×9档) + choice(8族训练+OCC-22解码) + noul×2(文本层信号)")
print("✓ battery 自检全部通过")
|