OracleProto:预测评估集
英文文档: [English doc]
GitHub 仓库: [MaYiding/OracleProto]
查看排行榜网站: [网站链接]
查看我们的论文: [arXiv]
一份以 SQLite 打包的评估集,共 80 道经人工精校、围绕现实事件的预测题,resolution date 介于 2026-03-12 与 2026-04-14 之间;与 GitHub 仓库 同步发布。题目行与字节稳定的提示重建配方共存于单个 forecast_eval_set_example.db 文件,其中两张表分别是 forecast_eval_set_example(80 行题目)与 dataset_metadata(配方)。
1. 数据集概览
| 字段 | 取值 |
|---|---|
| 发布日期 | 2026-04-29 |
| 行数 | 80 |
| Splits | train (80);单一 split,定位为留出评估集 |
| Resolution-date 范围 | 2026-03-12 → 2026-04-14 |
| 题目类型 | yes_no、binary_named、multiple_choice |
| Choice 类型 | single(恰一个正确字母)、multi(一个或多个正确字母) |
| 数据库文件 | forecast_eval_set_example.db(SQLite 3,约 52 KB) |
| 文件中的表 | forecast_eval_set_example(80 行)、dataset_metadata(1 行) |
| 协议 | MIT |
| 上游来源 | HuggingFace 预测题集(levels 1+2),原始 322 → 精校 80 |
类型分布
question_type |
choice_type |
行数 |
|---|---|---|
yes_no |
single |
37 |
binary_named |
single |
3 |
multiple_choice |
single |
32 |
multiple_choice |
multi |
8 |
| 合计 | 80 |
yes_no 是二元 Yes/No 题;binary_named 在两个命名实体之间二选一,例如两支球队、两名参赛者或两方对阵;multiple_choice 至少含三个带字母标签的选项,其中一个或多个为正确答案;选项列表中出现 None of the above 时它同样是合法答案。每行存储完整的选项标签字面值;字母 A 映射到 options[0],B 映射到 options[1],依此类推(§3.4 涵盖 Z 之后的标签情形)。
2. 文件
OracleProto/
├── forecast_eval_set_example.db # SQLite 数据库文件(数据集本体;约 52 KB)
├── forecast_eval_set_example.csv # 行表的 CSV 导出;80 行 + 表头(约 18 KB)
├── README.md # 本文件
├── LICENSE # MIT
└── .gitattributes # HF 标准二进制属性
数据集以单个 SQLite 文件(而非 Parquet 或 JSONL)发布,因为提示重建配方与逐行 provenance 与题目行共存于同一个文件(位于 dataset_metadata.features_json)。把行转换为 datasets.Dataset 的 loader 见 §6.3。
CSV 是 forecast_eval_set_example 行表的导出,不含 dataset_metadata,因此提示模板仅能从 SQLite 文件中获取。当下游流水线只需这 80 行(用于 pandas、电子表格或 grep 过滤)并自行重建提示时,使用 CSV。options 列保留为 JSON 编码的数组字符串,按 RFC 4180 转义。
3. 数据库 schema
两张表:forecast_eval_set_example 存有 80 行题目;dataset_metadata 存有规范配方。文件名取自主表。
3.1 表 forecast_eval_set_example(题目行)
CREATE TABLE forecast_eval_set_example (
id TEXT PRIMARY KEY,
choice_type TEXT NOT NULL CHECK (choice_type IN ('single','multi')),
question_type TEXT NOT NULL, -- yes_no | binary_named | multiple_choice
event TEXT NOT NULL, -- 待预测事件
options TEXT NOT NULL, -- 选项标签的 JSON 数组
answer TEXT NOT NULL, -- 规范化的正确答案,编码为字母
end_time TEXT NOT NULL -- 'YYYY-MM-DD'
);
CREATE INDEX idx_forecast_eval_set_example_choice_type ON forecast_eval_set_example(choice_type);
CREATE INDEX idx_forecast_eval_set_example_question_type ON forecast_eval_set_example(question_type);
CREATE INDEX idx_forecast_eval_set_example_end_time ON forecast_eval_set_example(end_time);
3.2 表 dataset_metadata(配方)
单行表,其 features_json blob 中存有提示模板、四种 output_format、outcomes-block 规则、agent role 字符串,以及 curation provenance。完整配方在 §5 中展开。
CREATE TABLE dataset_metadata (
dataset_name TEXT NOT NULL,
split_name TEXT NOT NULL,
table_name TEXT NOT NULL,
row_count INTEGER NOT NULL,
imported_at_utc TEXT NOT NULL,
features_json TEXT NOT NULL
);
3.3 列语义
| 列 | 类型 | 描述 |
|---|---|---|
id |
TEXT | 来自上游 HuggingFace 预测题集的稳定 source-side question ID;主连接键。 |
choice_type |
TEXT | 当且仅当一个字母正确时为 'single',可为一个或多个字母时为 'multi'。由 answer 中的字母个数推导。在 §5.4 中决定使用单选还是多选模板。 |
question_type |
TEXT | 取 yes_no、binary_named、multiple_choice 之一。决定渲染哪一种提示模板(§5)。 |
event |
TEXT | 待预测事件的自然语言描述;作者已编辑过,使时间锚定、单位与二元框架均明确。 |
options |
TEXT | 选项标签的 JSON 数组。yes_no 固定为 ["Yes","No"]。binary_named 是两个命名实体。multiple_choice 是选项标签的列表,字母由下标隐式确定(A=options[0], B=options[1], …)。 |
answer |
TEXT | 规范化的正确答案,编码为字母。yes_no 与 binary_named 为 'A' 或 'B'。multiple_choice 为按选项顺序排列、以逗号分隔的字母列表,例如 'A' 或 'A, B'。 |
end_time |
TEXT | resolution date,格式 YYYY-MM-DD。该列只存日历日期;GMT+8 的时区读法由提示模板(§5.2)在渲染时附上。如需更细粒度的 admissibility,可把每条 resolution 视为覆盖整个日历日。 |
3.4 字母到下标的编码
字母按 index = ord(letter) - ord('A') 映射到选项下标。超过 Z 之后(即 ≥27 个选项),标签沿以 A 起始的连续 ASCII 区间继续延伸:[、\、]、^、_、`、a、b、…。参考 renderer 会用反引号包裹任何非 A–Z 标签,使其在 markdown 渲染下保持可读。80 行中没有超过 26 个选项的题;之所以仍写入文档,是因为 framework 的 parser 支持该编码。
4. 行示例
{
"id": "699d9ffc098cca008728b6f0",
"choice_type": "single",
"question_type": "yes_no",
"event": "Will the US PCE annual inflation be greater than 2.9% in January 2026?",
"options": ["Yes", "No"],
"answer": "B",
"end_time": "2026-03-13"
}
{
"id": "69a2e39e5692ef005cdbf2d3",
"choice_type": "single",
"question_type": "binary_named",
"event": "Will US or Israel strike Iran first?",
"options": ["US", "Israel"],
"answer": "B",
"end_time": "2026-03-31"
}
{
"id": "6995b1073ea64b005b11f285",
"choice_type": "single",
"question_type": "multiple_choice",
"event": "Which men's basketball team will win the Big 12 Conference Championship tournament in the 2025-26 season?",
"options": ["Arizona", "Baylor", "Brigham Young University (BYU)",
"Houston", "Iowa State", "Kansas", "Kansas State"],
"answer": "A",
"end_time": "2026-03-14"
}
{
"id": "698f198bda7a8b006575444c",
"choice_type": "multi",
"question_type": "multiple_choice",
"event": "Which movies will win multiple Oscars? (2026)",
"options": ["One Battle After Another", "Sinners", "Frankenstein",
"KPop Demon Hunters", "F1", "Sentimental Value", "Hamnet",
"Marty Supreme", "The Secret Agent", "Avatar: Fire and Ash",
"Train Dreams", "Bugonia", "Blue Moon", "It Was Just An Accident"],
"answer": "A, B, C, D",
"end_time": "2026-03-15"
}
5. 提示重建(规范配方)
每一行通过 dataset_metadata.features_json.prompt_reconstruction 中的配方渲染为一条 user message。该配方字节稳定,是 OracleProto 评估器的事实来源;自行重建提示的下游用户应严格遵循,以保证结果可比。
5.1 静态片段
agent_role: "You are an agent that can predict future events."
guidance: "Do not use any other format. Do not refuse to make a prediction.
Do not say \"I cannot predict the future.\" You must make a clear
prediction based on the best data currently available, using the
box format specified above."
5.2 主模板
{agent_role} The event to be predicted: "{event} (resolved around {end_time} (GMT+8)).{outcomes_block}"
IMPORTANT: Your final answer MUST end with this exact format:
{output_format}
{guidance}
用户可见字符串中字面的 (GMT+8) 在渲染时给 resolution date 附上时区读法。
5.3 outcomes_block
yes_no 与 binary_named:为空,因为选项标签已嵌入 output_format。
multiple_choice:以一个换行符开头,随后每行一个选项,形式为 A. <label>,例如 \nA. Arizona\nB. Baylor\nC. Brigham Young University (BYU)\n…。若派生字母落在 A–Z 之外,则用反引号包裹该标签。
5.4 output_format(四选一,由 question_type × choice_type 决定)
yes_no:
Your task is to predict whether the event will occur based on your analysis.
Your prediction will be scored based on its accuracy. You will only receive points if your answer is correct.
Your final answer MUST end with this exact format:
\boxed{Yes} or \boxed{No}
**binary_named**(字面 <options[0]> 与 <options[1]> 替换为 options 中的两个命名实体):
Your task is to predict which of the two outcomes will occur based on your analysis.
Your prediction will be scored based on its accuracy. You will only receive points if your answer is correct.
Your final answer MUST end with this exact format:
\boxed{<options[0]>} or \boxed{<options[1]>}
multiple_choice 且 choice_type='single':
This is a SINGLE-ANSWER question: exactly ONE of the listed options is correct.
Your prediction will be scored on strict equality with the unique correct letter; choosing the wrong letter, or selecting more than one letter, scores zero.
Your final answer MUST end with this exact format:
the single correct letter inside the box, e.g. \boxed{A}.
Do NOT list more than one letter, even if you believe two outcomes are tied — pick the one you find most likely.
multiple_choice 且 choice_type='multi':
This is a MULTI-SELECT question: ONE OR MORE of the listed options can be correct.
Your prediction will be scored on strict equality with the FULL set of correct letters: any extra letter, any missing letter, or any wrong letter scores zero. You must include ALL correct options and NO incorrect options.
Your final answer MUST end with this exact format:
listing all correct option(s) you have identified, separated by commas, within the box.
For example: \boxed{A} for a single correct option, or \boxed{B, C} for multiple correct options.
5.5 答案解析
参考 parser(forecast_eval/parser.py::parse_answer)应用如下规则:
- 取模型回复中最后一个
\boxed{...}子串;其余视为 reasoning 或 scratchpad,忽略。 yes_no(不区分大小写):Yes→A,No→B。其余记为 unparsed。binary_named(不区分大小写):将盒内 payload 与options[0]或options[1]匹配。其余记为 unparsed。multiple_choice:按逗号与空白切分盒内 payload,校验每个 token 都是单字母,且每个字母都解析到合法的选项下标。越界字母或多字符 token 记为 unparsed。- 与从
answer解析出的规范字母集合做严格集合相等评分。缺失或 unparsed 的盒内答案记为parse_ok = 0,不视为 parser 错误;记录该状态后运行继续,不会中断。
复用 framework 的 parser 是跨实现获得 bit-identical 分数的最简单做法。
6. 加载数据集
6.1 使用原生 sqlite3(无额外依赖)
import sqlite3
import json
conn = sqlite3.connect("forecast_eval_set_example.db")
conn.row_factory = sqlite3.Row
# 读取题目行。
rows = conn.execute("SELECT * FROM forecast_eval_set_example").fetchall()
print(f"loaded {len(rows)} rows")
sample = dict(rows[0])
sample["options"] = json.loads(sample["options"]) # JSON 解码选项列表
print(sample)
# 读取规范化的提示重建配方。
meta_row = conn.execute("SELECT features_json FROM dataset_metadata").fetchone()
meta = json.loads(meta_row["features_json"])
prompt_template = meta["prompt_reconstruction"]["prompt_template"]
print(prompt_template)
6.2 使用 huggingface_hub
from huggingface_hub import hf_hub_download
import sqlite3, json
db_path = hf_hub_download(
repo_id="MaYiding/OracleProto",
filename="forecast_eval_set_example.db",
repo_type="dataset",
)
conn = sqlite3.connect(db_path)
rows = conn.execute("SELECT * FROM forecast_eval_set_example").fetchall()
6.3 转换为 datasets.Dataset
import sqlite3, json
from datasets import Dataset
conn = sqlite3.connect("forecast_eval_set_example.db")
cur = conn.execute("SELECT * FROM forecast_eval_set_example")
cols = [c[0] for c in cur.description]
def _row(r):
d = dict(zip(cols, r))
d["options"] = json.loads(d["options"]) # list[str]
d["answer_letters"] = [
s.strip() for s in d["answer"].split(",") if s.strip()
] # list[str]
return d
ds = Dataset.from_list([_row(r) for r in cur.fetchall()])
print(ds)
print(ds[0])
6.4 渲染提示(最小实现,遵从规范配方)
def render_prompt(row, meta):
rcp = meta["prompt_reconstruction"]
options = row["options"]
qt, ct = row["question_type"], row["choice_type"]
if qt == "yes_no":
outcomes_block = ""
out_fmt = rcp["yes_no_output_format"]
elif qt == "binary_named":
outcomes_block = ""
out_fmt = (
rcp["binary_named_output_format"]
.replace("<options[0]>", options[0])
.replace("<options[1]>", options[1])
)
elif qt == "multiple_choice":
outcomes_block = "\n" + "\n".join(
f"{chr(ord('A') + i)}. {label}" for i, label in enumerate(options)
)
key = (
"multiple_choice_single_output_format" if ct == "single"
else "multiple_choice_multi_output_format"
)
out_fmt = rcp[key]
else:
raise ValueError(qt)
return rcp["prompt_template"].format(
agent_role=rcp["agent_role"],
event=row["event"],
end_time=row["end_time"],
outcomes_block=outcomes_block,
output_format=out_fmt,
guidance=rcp["guidance"],
)
完整参考 renderer(含 >26 选项的反引号规则与可选的 reflection / belief-elicitation 尾部)位于 forecast_eval/prompts.py;复用它即可获得 byte-identical 提示。
6.5 使用预生成的 CSV(标准库 csv,不含提示模板)
import csv, json
with open("forecast_eval_set_example.csv", encoding="utf-8", newline="") as f:
rows = [
{**r, "options": json.loads(r["options"])}
for r in csv.DictReader(f)
]
print(f"loaded {len(rows)} rows; first event: {rows[0]['event']!r}")
CSV 路径完全绕过 dataset_metadata。要把行与提示模板配对,要么按 §5 手工渲染,要么切换回 §6.1 的 SQLite 路径。
7. 推荐评估协议
将本数据集与 OracleProto 评估 harness 配套使用,后者在朴素的「提示+评分」循环之上叠加信息边界纪律。五条具体建议:
为每个模型声明 knowledge cutoff $\kappa_M$。 题目 $i$ 对模型 $M$ 而言 admissible 的条件是 $\kappa_M \le \chi_i < \tau_i$,其中 $\chi_i$ 是题目的 prediction cutoff,$\tau_i$ 是其 resolution date。Inadmissible 题目在上游过滤,不计入模型错误。未声明 cutoff 的模型无法与已声明的模型公平比较。
对任何 retrieval 或 browsing 工具做时间掩蔽。 当 harness 允许模型发起 web 搜索时,将搜索侧
end_date锁定到 $\chi_i + \delta$ 并采用保守的 offset;OracleProto 默认 $\delta = -1$ 天。该屏障 (L2) 背后的机制记录在 framework 的 DESIGN 与 FRAME 文档中。运行独立的 retrieval-content 审计员。 每条召回片段交由独立 LLM 审计员判断是否泄露 resolution。这是 framework 威胁模型中的 L3 屏障。
禁用 provider-native browsing。 OracleProto 在三层上拒收以
:online等 hosted-browsing 变体结尾的 model slug:config 校验、on-the-wire client 与 detector client。这是 L4 屏障,也是任何一次计费 LLM 调用离开进程前必须通过的最终检查。以字母集合上的严格集合相等评分,参 §5.5。当模型按 framework 的 belief-elicitation 协议额外输出
<belief>{ ... }</belief>JSON 块时,可选启用概率 calibration 指标(Brier、NLL、ECE、Murphy 分解);schema 见forecast_eval/prompts.py::BELIEF_PROTOCOL。
未启用 OracleProto harness 时,应将所得数字视为预测能力的上界:任何能浏览开放 web、或训练截止越过题目 end_time 的模型都可能记忆了答案。数据集使得 admissibility 检查成为可能,但它本身并不强制执行。
8. Provenance 与 curation
- 来源。 上游 HuggingFace 预测题集,限制在 levels 1+2(上游难度带中较容易的两档)。从原始集合采集到 322 道候选题。
- Curation 流水线(5 pass)。
- Source-side 坏行剔除与列扁平化。
end_time/ 答案编码 / 选项标签规范化:end_time归一为YYYY-MM-DD日历日期;Yes/No映射为A/B;选项标签去除残余 markdown。- 322 → 200 → 100 → 80 的下采样,并伴有 placeholder 移除、去重与歧义审计。
- 终轮 HIGH+MEDIUM 歧义修复:重写 4 行,使时间锚定、单位与二元框架均明确。
- 对一道 S&P 500 multi-select 真值集做 CRITICAL 修复,使其满足选项阶梯隐含的单调阈值逻辑。
- 验证。 全部 80 条 ground truth 通过 parser 往返做端到端验证(rendered prompt 经 parse 后再编码回规范字母集合)。最终计数:剩余 0 critical / 0 high / 0 medium 歧义问题。
9. 用途与局限
9.1 适用场景
- LLM 与 LLM 智能体的预测基准,特别是结合参数化知识与时间掩蔽 web retrieval 的工具型智能体。
- 预测 harness 的复现性试验台。
dataset_metadata表使每条提示字节稳定;与 OracleProto framework 配套使用时,可得到一个运行单元,其评分工件在配置匹配时 bit-identical。 - 校准与 proper-scoring 研究。 80 行规模足够小,使逐题分析(信念演化、来源归因、calibration 图)仍在可处理范围内。
9.2 不适用场景
- 长时程预测。 所有 resolution 落在一个月窗口(2026-03-12 → 2026-04-14);该集合不代表跨季度或跨年度预测。
- 开放生成。 每题都有封闭答案集,因此并非生成基准。
10. License
按 MIT License 发布(见 LICENSE)。上游题目源自公开 HuggingFace 预测集;本版本中的 curation 工作、schema、提示重建配方与答案编码均为本项目的贡献。
11. 联系与合作
如有代码使用、数据集构建、复现问题等,欢迎直接联系项目开发者:
如需联合研究、数据与评测基准共建、论文合作等,请联系课题负责人:
- 黄凯博(通讯作者):huangkaibo@bupt.edu.cn
- 杨忠良(通讯作者):yangzl@bupt.edu.cn
12. 论文
查看我们的论文:arXiv
13. 引用
如果您在研究中使用了本项目,请引用我们的论文:
@article{OracleProto,
title={OracleProto: A Reproducible Framework for Benchmarking LLM Native Forecasting via Knowledge Cutoff and Temporal Masking},
author={Yiding Ma, Chengyun Ruan, Kaibo Huang, Zhongliang Yang, Linna Zhou},
journal={arXiv preprint arXiv:2605.03762},
year={2026}
}