code_pipeline/data_construction/config.py

import os
from pathlib import Path

# === Paths ===
PROJECT_ROOT = Path(__file__).resolve().parent.parent.parent
DATA_DIR = PROJECT_ROOT / "data" / "custom_history_data"
OUTPUT_DIR = PROJECT_ROOT / "output"
ERROR_DIR = OUTPUT_DIR / "errors"

# === LLM Configuration ===
LLM_PROVIDER = os.getenv("LLM_PROVIDER", "mify")  # mify | openai | vllm
LLM_BASE_URL = os.getenv("LLM_BASE_URL", None)
LLM_API_KEY = os.getenv("LLM_API_KEY", None)

# 网关 api.llm.mioffice.cn 可用: "ppio/pa/gpt-5.5" / "glm-5.2" (zhipuai) / "gemini-3.1-pro-preview-pt" (vertex_ai)
LLM_MODEL_GENERATION = os.getenv("LLM_MODEL_GENERATION", "ppio/pa/gpt-5.5")
LLM_MODEL_PREDICTION = os.getenv("LLM_MODEL_PREDICTION", "ppio/pa/gpt-5.5")
LLM_MODEL_TEACHER = os.getenv("LLM_MODEL_TEACHER", "ppio/pa/gpt-5.5")

# === Concurrency ===
# 数据生成并行度. 通过 ThreadPoolExecutor 并行调用 LLM API.
# 8 是经验值, 上调到 16-32 在 API rate limit 允许下能更快, 但容易触发限流.
MAX_WORKERS = int(os.getenv("MAX_WORKERS", "32"))
# 质检并行度 (三教师投票). Claude API 并发限频严重, 需要比生成低.
MAX_WORKERS_JUDGE = int(os.getenv("MAX_WORKERS_JUDGE", "4"))

# === Data Scale ===
NUM_USERS = 500                  # 接着 user_170 (u0001~u0200) 增量补量到 u0600
QA_PER_USER_AVG = 10
TOTAL_QA_TARGET = 4000           # 全局配额上限 (须 >= 旧QA + 新增预期, 否则旧QA占满配额新用户不生成)
NUM_FILLERS = 29000
FILLER_TO_EVIDENCE_RATIO = 5     # 对齐 user_170 (长版 ~111k): 每条 evidence 配 5 条 filler
SESSIONS_PER_USER_EVIDENCE = 10
SESSIONS_PER_USER_TANGENTIAL = 6  # 对齐 user_170

# === Query Type Distribution ===
QUERY_TYPES = [
    "single_hop",
    "knowledge_update",
    "multi_session_synthesis",
    "two_hop",
    "temp_reasoning_implicit",
    "temp_reasoning_explicit",
    "implicit_preference",
    "assistant_previnfo",
]

# v2 second-pass 配额 (600 用户): 易类配额上调以吸纳新用户产能, 难类保持高位
# 利用 v2 first-pass 实测通过率 (single 100%, assist 94%, two_hop 91%, multi 23%,
# pref 53%, ku 7.6%, temp_explicit 5.8%, temp_implicit 0%) 反推:
QA_PER_TYPE = {
    "single_hop":              500,   # 全局上限, 含已有旧QA; 设高只是不卡新用户, 填不满不报错
    "assistant_previnfo":      500,
    "implicit_preference":     500,
    "knowledge_update":        500,
    "two_hop":                 500,
    "multi_session_synthesis": 500,
    "temp_reasoning_implicit": 500,
    "temp_reasoning_explicit": 500,
}
assert sum(QA_PER_TYPE.values()) == TOTAL_QA_TARGET, \
    f"QA_PER_TYPE sum {sum(QA_PER_TYPE.values())} != TOTAL_QA_TARGET {TOTAL_QA_TARGET}"
assert set(QA_PER_TYPE.keys()) == set(QUERY_TYPES), "QA_PER_TYPE keys must match QUERY_TYPES"

# 每个用户最多生成多少条 QA. v2: 6 -> 8 (3000/400=7.5, 取上界 8)
PER_USER_QA_CAP = int(os.getenv("PER_USER_QA_CAP", "14"))

# === Generation Parameters ===
PRED_TEMPERATURE = 0.7
PRED_NUM_SAMPLES = 3
TEACHER_NUM_VOTES = 3
GENERATION_TEMPERATURE = 0.9
MAX_RETRIES = 10

# Output token budget. gpt-5.5 是 reasoning 模型, max_completion_tokens
# 包含 hidden reasoning tokens, 所以要比纯 chat 模型留更多余量, 否则 JSON
# 容易被截断 (出现 "JSON parse failed" 警告).
MAX_TOKENS_GENERATION = 8192          # 单 session / profile / QA pair
MAX_TOKENS_GENERATION_DUAL = 16384    # knowledge_update / two_hop 一次返回两个 session

# === Inline Check Pipeline ===
# 默认开启每步生成完后立即跑 check + 过滤, 下游只看 strict 通过的子集.
# 可通过 INLINE_CHECK=0 关闭, 走旧的"全量生成 + 末尾批量 check" 流程.
INLINE_CHECK = os.getenv("INLINE_CHECK", "1") == "1"

# === Filtering Thresholds ===
VOTE_AGREEMENT_THRESHOLD = 0.67
CONFIDENCE_THRESHOLD = 0.60
LOW_CONF_WEIGHT = 0.3

# === Retrieval ===
BM25_TOP_K = 10

# === Output Files ===
OUTPUT_A1 = OUTPUT_DIR / "a1_user_profiles.json"
OUTPUT_A1_STRICT = OUTPUT_DIR / "a1_user_profiles_strict.json"
OUTPUT_A2 = OUTPUT_DIR / "a2_user_sessions.json"
OUTPUT_A2_STRICT = OUTPUT_DIR / "a2_user_sessions_strict.json"
OUTPUT_A3 = OUTPUT_DIR / "a3_qa_pairs.json"
OUTPUT_A3_STRICT = OUTPUT_DIR / "a3_qa_pairs_strict.json"
OUTPUT_A4 = OUTPUT_DIR / "a4_filler_sessions.json"
OUTPUT_A5 = OUTPUT_DIR / "a5_user_data.json"
OUTPUT_B1 = OUTPUT_DIR / "b1_predictions.json"
OUTPUT_B2 = OUTPUT_DIR / "b2_ms_labels.json"
OUTPUT_B3 = OUTPUT_DIR / "b3_rewrite_queries.json"
OUTPUT_B4 = OUTPUT_DIR / "b4_validated.json"
OUTPUT_B5 = OUTPUT_DIR / "metamem_train.json"