Datasets:

sselaine27
/

benchmark-research

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
dataset_name: string
version: string
status: string
license: string
task_files: struct<A1: struct<file: string, status: string, prompt_template: string, builder: string>, A2-F: str (... 599 chars omitted)
  child 0, A1: struct<file: string, status: string, prompt_template: string, builder: string>
      child 0, file: string
      child 1, status: string
      child 2, prompt_template: string
      child 3, builder: string
  child 1, A2-F: struct<file: string, status: string, prompt_template: string, builder: string>
      child 0, file: string
      child 1, status: string
      child 2, prompt_template: string
      child 3, builder: string
  child 2, A2-T: struct<file: string, status: string, prompt_template: string, builder: string>
      child 0, file: string
      child 1, status: string
      child 2, prompt_template: string
      child 3, builder: string
  child 3, A2-H: struct<file: string, status: string, prompt_template: string, builder: string>
      child 0, file: string
      child 1, status: string
      child 2, prompt_template: string
      child 3, builder: string
  child 4, B: struct<file: string, template_file: string, status: string, prompt_template: string, builder: string (... 31 chars omitted)
      child 0, file: string
      child 1, template_file: string
      child 2, status: string
      child 3, prompt_template: string
      child 4, builder: string
      child 5, variants: list<item: string>
          child 0, item: string
  child 5, C: struct<f
...
 child 2, prompt_template: string
      child 3, builder: string
  child 6, D: struct<file: string, status: string, prompt_template: string>
      child 0, file: string
      child 1, status: string
      child 2, prompt_template: string
  child 7, E: struct<file: string, status: string, prompt_template: string>
      child 0, file: string
      child 1, status: string
      child 2, prompt_template: string
data_files: struct<a1_price_snapshots.csv: string, a2_price_series.csv: string, a2_fundamentals_snapshot.csv: st (... 93 chars omitted)
  child 0, a1_price_snapshots.csv: string
  child 1, a2_price_series.csv: string
  child 2, a2_fundamentals_snapshot.csv: string
  child 3, a2_cohorts_manual.csv: string
  child 4, c_financial_snapshots.csv: string
  child 5, b_events.csv: string
counts: struct<A1_ready: int64, A2-F_ready: int64, A2-T_ready: int64, A2-H_ready: int64, B_ready: int64, C_r (... 97 chars omitted)
  child 0, A1_ready: int64
  child 1, A2-F_ready: int64
  child 2, A2-T_ready: int64
  child 3, A2-H_ready: int64
  child 4, B_ready: int64
  child 5, C_ready: int64
  child 6, D_ready: int64
  child 7, E_ready: int64
  child 8, template_records: int64
  child 9, total_ready_records: int64
validation: struct<script: string, command: string>
  child 0, script: string
  child 1, command: string
notes: string
agent: string
api_key_env: string
temperature: double
categories: list<item: string>
  child 0, item: string
max_tokens: int64
model: string
timeout_seconds: double
to
{'agent': Value('string'), 'model': Value('string'), 'api_key_env': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64'), 'timeout_seconds': Value('float64'), 'categories': List(Value('string')), 'notes': Value('string')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 478, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2818, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              dataset_name: string
              version: string
              status: string
              license: string
              task_files: struct<A1: struct<file: string, status: string, prompt_template: string, builder: string>, A2-F: str (... 599 chars omitted)
                child 0, A1: struct<file: string, status: string, prompt_template: string, builder: string>
                    child 0, file: string
                    child 1, status: string
                    child 2, prompt_template: string
                    child 3, builder: string
                child 1, A2-F: struct<file: string, status: string, prompt_template: string, builder: string>
                    child 0, file: string
                    child 1, status: string
                    child 2, prompt_template: string
                    child 3, builder: string
                child 2, A2-T: struct<file: string, status: string, prompt_template: string, builder: string>
                    child 0, file: string
                    child 1, status: string
                    child 2, prompt_template: string
                    child 3, builder: string
                child 3, A2-H: struct<file: string, status: string, prompt_template: string, builder: string>
                    child 0, file: string
                    child 1, status: string
                    child 2, prompt_template: string
                    child 3, builder: string
                child 4, B: struct<file: string, template_file: string, status: string, prompt_template: string, builder: string (... 31 chars omitted)
                    child 0, file: string
                    child 1, template_file: string
                    child 2, status: string
                    child 3, prompt_template: string
                    child 4, builder: string
                    child 5, variants: list<item: string>
                        child 0, item: string
                child 5, C: struct<f
              ...
               child 2, prompt_template: string
                    child 3, builder: string
                child 6, D: struct<file: string, status: string, prompt_template: string>
                    child 0, file: string
                    child 1, status: string
                    child 2, prompt_template: string
                child 7, E: struct<file: string, status: string, prompt_template: string>
                    child 0, file: string
                    child 1, status: string
                    child 2, prompt_template: string
              data_files: struct<a1_price_snapshots.csv: string, a2_price_series.csv: string, a2_fundamentals_snapshot.csv: st (... 93 chars omitted)
                child 0, a1_price_snapshots.csv: string
                child 1, a2_price_series.csv: string
                child 2, a2_fundamentals_snapshot.csv: string
                child 3, a2_cohorts_manual.csv: string
                child 4, c_financial_snapshots.csv: string
                child 5, b_events.csv: string
              counts: struct<A1_ready: int64, A2-F_ready: int64, A2-T_ready: int64, A2-H_ready: int64, B_ready: int64, C_r (... 97 chars omitted)
                child 0, A1_ready: int64
                child 1, A2-F_ready: int64
                child 2, A2-T_ready: int64
                child 3, A2-H_ready: int64
                child 4, B_ready: int64
                child 5, C_ready: int64
                child 6, D_ready: int64
                child 7, E_ready: int64
                child 8, template_records: int64
                child 9, total_ready_records: int64
              validation: struct<script: string, command: string>
                child 0, script: string
                child 1, command: string
              notes: string
              agent: string
              api_key_env: string
              temperature: double
              categories: list<item: string>
                child 0, item: string
              max_tokens: int64
              model: string
              timeout_seconds: double
              to
              {'agent': Value('string'), 'model': Value('string'), 'api_key_env': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64'), 'timeout_seconds': Value('float64'), 'categories': List(Value('string')), 'notes': Value('string')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Benchmark Research

面向金融 Deep Research Agent 的评测数据集仓库。

版本与状态

v0.2.0 — dataset release candidate

类别	Ready 条数	文件
A1	20	`seeds/a1_valuation.jsonl`
A2-F	2	`seeds/a2_fundamentals.jsonl`
A2-T	2	`seeds/a2_technical.jsonl`
A2-H	2	`seeds/a2_hybrid.jsonl`
B	0	`seeds/b_event.jsonl`（builder 已接通，数据待填充）
C	76	`seeds/c_financial_metric.jsonl`
D	6	`seeds/d_counterfactual.jsonl`
E	5	`seeds/e_formula.jsonl`
合计	113	—
Template 样例	10	`seeds/templates/*.jsonl`

本版本完成：

统一 JSONL seed schema 与 8 个 prompt 模板
A1 / A2 / C / D / E 的 ready seeds 与本地 builder + runner 闭环
B 的 builder / parser / evaluator 已接通；data/b_events.csv 当前为空
结构校验脚本覆盖全部 ready seed 文件
运行产物目录 results/ 已 gitignore，仓库内仅保留占位

详细任务说明见 docs/task_cards.md。机器可读元数据见 manifest.json。

任务总览

ID	任务	说明	当前状态
A1	单股估值区间预测	bull/base/bear 三情景 + 修复期	ready（20）
A2-F	行业横截面排序（基本面）	仅财务指标	ready（2）
A2-T	行业横截面排序（技术面）	仅预计算技术指标	ready（2）
A2-H	行业横截面排序（混合）	基本面 + 技术	ready（2）
B	事件驱动方向预测	earnings / macro	builder ready, data empty
C	财务指标前向预测	非直觉指标数值推断	ready（76）
D	反事实事件注入	虚构新闻 + 逻辑方向判分	ready（6）
E	多步金融公式计算	精确数值 + 公式选择	ready（5）

文件结构

benchmark-research/
├── README.md
├── manifest.json
├── .gitignore
├── data/                          # 输入 CSV（构建 seeds 用）
│   ├── a1_price_snapshots.csv
│   ├── a2_price_series.csv
│   ├── a2_fundamentals_snapshot.csv
│   ├── a2_cohorts_manual.csv
│   ├── c_financial_snapshots.csv
│   └── b_events.csv               # 当前为空
├── scripts/
│   └── validate.py
├── docs/
│   ├── schema.md
│   └── task_cards.md
├── prompts/                       # 8 个 prompt 模板
├── seeds/
│   ├── a1_valuation.jsonl         # ready
│   ├── a2_fundamentals.jsonl      # ready
│   ├── a2_technical.jsonl         # ready
│   ├── a2_hybrid.jsonl            # ready
│   ├── b_event.jsonl              # ready（0 条）
│   ├── c_financial_metric.jsonl   # ready
│   ├── d_counterfactual.jsonl     # ready
│   ├── e_formula.jsonl            # ready
│   └── templates/                 # schema / prompt 对齐样例（10 条）
├── src/
│   ├── builders/                  # CSV → JSONL
│   ├── parsers/                   # Agent 输出解析
│   ├── evaluators/                # 指标计算
│   ├── agents/                    # mock / HF baseline
│   └── run_benchmark.py
└── results/                       # 本地运行输出（gitignore）
    └── .gitkeep

JSONL Record Format

每行一个 JSON object，统一字段：

字段	说明
`task_id`	全局唯一 ID
`category`	A1 / A2 / B / C / D / E
`variant`	子变体（如 F/T/H、earnings）
`time_band`	T1 / T2 / T3 / null
`status`	`template` / `ready` / `validated`
`seed`	结构化输入
`prompt`	已渲染的完整 prompt
`expected_output`	输出 JSON schema 说明
`ground_truth`	评测标签（template 可为 null）
`metadata`	元信息（含 `is_template`）

完整 schema 说明见 docs/schema.md。

Agent 接口

Agent 被视为黑盒：

def run_agent(prompt: str) -> str:
    ...

本地加载示例

import json

def load_jsonl(path: str) -> list[dict]:
    with open(path, encoding="utf-8") as f:
        return [json.loads(line) for line in f if line.strip()]

# Ready seeds
a1 = load_jsonl("seeds/a1_valuation.jsonl")
c_tasks = load_jsonl("seeds/c_financial_metric.jsonl")

print(a1[0]["prompt"])
print(c_tasks[0]["ground_truth"])

# Template 样例
templates = load_jsonl("seeds/templates/a1_valuation_template.jsonl")
assert templates[0]["metadata"]["is_template"] is True

结构校验

python scripts/validate.py

校验项：必需文件存在、JSONL 格式、统一字段、template/ready 状态、metadata.is_template、ready 文件无 PLACEHOLDER、task_id 唯一性、D 题价格序列长度、A2 stock_list 至少 6 只等。

运行 Benchmark

支持 mock（本地闭环 smoke test）与 hf（Hugging Face Inference 单轮 baseline）两种 agent。所有类别 A1/A2/B/C/D/E 均已接入 parser 与 evaluator。

# Mock smoke test（任意 ready seed 文件）
python -m src.run_benchmark \
  --seed seeds/c_financial_metric.jsonl \
  --agent mock \
  --output results/mock_c

# Hugging Face baseline（需设置 token 环境变量）
export HF_TOKEN=your_token_here

python -m src.run_benchmark \
  --seed seeds/e_formula.jsonl \
  --agent hf \
  --model Qwen/Qwen3-8B \
  --api-key-env HF_TOKEN \
  --output results/hf_e \
  --limit 2

说明：

--limit N 仅用于 smoke test；正式跑完整集时去掉
不要把 token 写入代码或配置文件
输出写入 --output 目录：predictions.jsonl、metrics_summary.json、run_config.json
results/ 已在 .gitignore 中，需本地重新生成

各类别 Agent 输出格式

类别	输出 JSON
A1	`bull`, `base`, `bear`, `reversion_horizon`
A2	`["code_rank_1", "code_rank_2", ...]`（排名数组）
B	`direction`（`up`/`down`）, `probability_up`
C	`predicted_value`
D	`logic_direction`（`up`/`down`/`neutral`）
E	`formula_choice`, `computed_value`

从 CSV 构建 Seeds

A1

python -m src.builders.a1_from_csv_builder \
  --csv data/a1_price_snapshots.csv \
  --output seeds/a1_valuation.jsonl

A2-F

python -m src.builders.a2_fundamentals_from_csv_builder \
  --fundamentals-csv data/a2_fundamentals_snapshot.csv \
  --output seeds/a2_fundamentals.jsonl

A2-T / A2-H

python -m src.builders.a2_technicals_from_csv_builder \
  --price-series-csv data/a2_price_series.csv \
  --output seeds/a2_technical.jsonl

python -m src.builders.a2_hybrid_from_csv_builder \
  --fundamentals-csv data/a2_fundamentals_snapshot.csv \
  --price-series-csv data/a2_price_series.csv \
  --output seeds/a2_hybrid.jsonl

技术指标口径（写死）：

输入：cutoff 前最近 40 个交易日收盘价，升序，P0 为最后一根
rsi_14：14 期简单平均 RSI（非 Wilder smoothing）
macd_histogram：EMA12/EMA26，DIF 的 9 日 EMA 为 DEA，histogram = DIF - DEA
momentum_20d：(P0 - P_-20) / P_-20
bollinger_zscore：(P0 - MA20) / STD20（ddof=1）；STD20 == 0 时为 null

C

python -m src.builders.c_financial_metric_from_csv_builder \
  --csv data/c_financial_snapshots.csv \
  --output seeds/c_financial_metric.jsonl

B

python -m src.builders.b_event_from_csv_builder \
  --csv data/b_events.csv \
  --output seeds/b_event.jsonl

当前 data/b_events.csv 仅有表头，生成 0 条 ready。支持 earnings 与 macro 子类；policy 仅在 template 中作 schema 样例。

数据文件（`data/`）

文件	用途
`a1_price_snapshots.csv`	A1 单股估值快照
`a2_price_series.csv`	A2-T/H 价格序列与收益
`a2_fundamentals_snapshot.csv`	A2-F/H 基本面估值快照
`a2_cohorts_manual.csv`	A2-F cohort 定义
`c_financial_snapshots.csv`	C 财务指标快照
`b_events.csv`	B 事件驱动方向预测（当前为空）

估值快照日期不匹配（prototype）

当前 a2_fundamentals_snapshot.csv 可能仅有较晚日期（如 2026-03-22），而 cohort cutoff_date 更早（如 2025-06-06）。A2-F/H builder 匹配规则：

优先：trading_day <= cutoff_date 的最近一条（on_or_before_cutoff）
Fallback：全表最近一条（prototype_fallback_nearest），记录在 metadata

正式数据补齐 cutoff 当日或之前的历史估值后，fallback 将自动不再触发。

License

CC-BY-4.0

后续计划

填充 data/b_events.csv，生成 B ready seeds
扩充 A1/A2/D/E 规模
补齐 A2 fundamentals 历史快照，去除 prototype fallback
对 ready seeds 执行 validated 校验流程
发布公开版（视情况隐藏 ground_truth 中的未来价格）

Downloads last month: 68

Total file size:

886 kB