Yilinss27 commited on
Commit
93708ff
·
verified ·
1 Parent(s): 21f0f4d

Initial upload of FinForesight dataset

Browse files
README.md ADDED
@@ -0,0 +1,164 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-4.0
3
+ language:
4
+ - zh
5
+ pretty_name: Benchmark Research
6
+ tags:
7
+ - finance
8
+ - benchmark
9
+ - deep-research
10
+ - agent-evaluation
11
+ size_categories:
12
+ - n<1K
13
+ configs:
14
+ - config_name: val
15
+ data_files:
16
+ - split: train
17
+ path: data/01_val/train.jsonl
18
+ default: true
19
+ - config_name: rank_f
20
+ data_files:
21
+ - split: train
22
+ path: data/02_rank_f/train.jsonl
23
+ - config_name: rank_t
24
+ data_files:
25
+ - split: train
26
+ path: data/03_rank_t/train.jsonl
27
+ - config_name: rank_h
28
+ data_files:
29
+ - split: train
30
+ path: data/04_rank_h/train.jsonl
31
+ - config_name: evt
32
+ data_files:
33
+ - split: train
34
+ path: data/05_evt/train.jsonl
35
+ - config_name: fund
36
+ data_files:
37
+ - split: train
38
+ path: data/06_fund/train.jsonl
39
+ - config_name: cf
40
+ data_files:
41
+ - split: train
42
+ path: data/07_cf/train.jsonl
43
+ - config_name: calc
44
+ data_files:
45
+ - split: train
46
+ path: data/08_calc/train.jsonl
47
+ ---
48
+
49
+ # Benchmark Research
50
+
51
+ 面向金融 Deep Research Agent 的评测数据集(**仅数据**)。
52
+
53
+ 代码、builder、prompt 模板与评测脚本在独立代码仓库维护;本 Hub 仓库只发布标准化 ready seeds,支持按子集按需加载。
54
+
55
+ 任务命名:`val` / `rank_*` / `evt` / `fund` / `cf` / `calc`
56
+ (对应 Valuation / Return Ranking / Event Direction / Fundamentals / Counterfactual / Calculation)。
57
+
58
+ ## 快速开始
59
+
60
+ ```python
61
+ from datasets import load_dataset
62
+
63
+ # 按题型子集加载(推荐)
64
+ val = load_dataset("sselaine27/benchmark-research", "val")
65
+ fund = load_dataset("sselaine27/benchmark-research", "fund")
66
+ cf = load_dataset("sselaine27/benchmark-research", "cf")
67
+
68
+ print(val["train"][0]["task_id"])
69
+ print(val["train"][0]["prompt"][:200])
70
+
71
+ # 按时间层筛选
72
+ t2 = val["train"].filter(lambda x: x["time_band"] == "T2")
73
+ ```
74
+
75
+ 可用 config:`val` / `rank_f` / `rank_t` / `rank_h` / `evt` / `fund` / `cf` / `calc`
76
+ 默认 config 为 `val`。
77
+
78
+ ### EVT / FUND 扰动题
79
+
80
+ `evt` 与 `fund` 中包含扰动题:每条都由同类别的一道原题改写而来,问的是同一个问题,
81
+ `ground_truth` 与原题完全一致,用于测量同一问题在改写后的作答稳定性。
82
+
83
+ - `variant` 为 `perturb_layout`(版式重排)/ `perturb_paraphrase`(指令同义改写)/
84
+ `perturb_numeric`(披露数字按同一因子缩放 ±1%);其余 `variant` 为原题。
85
+ - `metadata.perturbation.base_task_id` 指向原题,`metadata.perturbation.perturbation_type`
86
+ 给出扰动类型。
87
+ - 只评原题时按 `variant` 过滤即可:
88
+
89
+ ```python
90
+ evt = load_dataset("sselaine27/benchmark-research", "evt")
91
+ base = evt["train"].filter(lambda x: not x["variant"].startswith("perturb_"))
92
+ ```
93
+
94
+ ## 规模(v0.9.0)
95
+
96
+ | config | 任务 | ready 条数 |
97
+ |--------|------|-----------|
98
+ | `val` | Valuation (VAL) | 244 |
99
+ | `rank_f` | Return Ranking F (RANK-F) | 112 |
100
+ | `rank_t` | Return Ranking T (RANK-T) | 161 |
101
+ | `rank_h` | Return Ranking H (RANK-H) | 110 |
102
+ | `evt` | Event Direction (EVT) | 228 |
103
+ | `fund` | Fundamentals (FUND) | 384 |
104
+ | `cf` | Counterfactual (CF) | 30 |
105
+ | `calc` | Calculation (CALC) | 30 |
106
+ | **合计** | | **1299** |
107
+
108
+ ### 实验身份(paper 主口径)
109
+
110
+ - `backbone_model = gpt-4.1`
111
+ - `backbone_training_cutoff = 2024-06-30`(OpenAI 仅声明 “June 2024”;本库按月末保守处理)
112
+ - `guard_days = 30` → T1 outcome ≤ `2024-05-31`,T2 origin ≥ `2024-07-30`
113
+ - `experiment_as_of = 2026-08-17`
114
+
115
+ **不兼容声明:** 旧 Hub pin(339 题 / `model_training_cutoff=2024-06-01`)为 legacy identity,不得与 v0.9 paper_band / C1 表混比。
116
+
117
+ ## 字段说明
118
+
119
+ 每行一条评测题(JSON object):
120
+
121
+ | 字段 | 说明 |
122
+ |------|------|
123
+ | `task_id` | 全局唯一 ID |
124
+ | `category` | VAL / RANK / EVT / FUND / CF / CALC |
125
+ | `variant` | 子变体(如 F/T/H、earnings) |
126
+ | `cutoff_date` | 用于时间分层的截止日期 |
127
+ | `time_band` | T1 / T2 / T3(legacy;论文请用 `paper_band`) |
128
+ | `forecast_origin` | 预测时点 |
129
+ | `outcome_available_at` | 结果可用日;结合 `outcome_available_at_source` 判断是否为观测值或估计值 |
130
+ | `paper_band` | GPT-4.1 论文口径 T1 / T2 / T3 / CF / CALC |
131
+ | `status` | ready |
132
+ | `seed` | 结构化输入 |
133
+ | `prompt` | 已渲染完整 prompt |
134
+ | `expected_output` | 输出 schema 说明 |
135
+ | `ground_truth` | 评测标签 |
136
+ | `metadata` | 元信息 |
137
+
138
+ ## 时间分层
139
+
140
+ 本数据集按可复算规则标注 legacy `time_band`(仅按 `cutoff_date`):
141
+
142
+ - `T1`:`cutoff_date <= model_training_cutoff`,可能存在预训练数据泄露
143
+ - `T2`:`model_training_cutoff < cutoff_date < reference_current_date`
144
+ - `T3`:`cutoff_date >= reference_current_date`,用于未来预测能力测试
145
+
146
+ 默认 legacy 参数(与 paper 身份对齐 cutoff 月,但规则不同):
147
+
148
+ - `model_training_cutoff = 2024-06-30`
149
+ - `reference_current_date = 2026-08-17`
150
+
151
+ 论文实验请使用侧车 `paper_band`(见代码仓 `docs/schema.md` / `docs/c1_pairing_contract.md`),不要直接用 legacy `time_band` 做 C1。
152
+
153
+ ## 仓库��界
154
+
155
+ 本仓库**只包含**:
156
+
157
+ - `README.md`(dataset card)
158
+ - `data/<config>/train.jsonl`(ready seeds)
159
+
160
+ **不包含**代码、scripts、prompts、原始构建 CSV、运行结果等。
161
+
162
+ ## License
163
+
164
+ [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)
data/01_val/train.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
data/02_rank_f/train.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
data/03_rank_t/train.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
data/04_rank_h/train.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
data/05_evt/train.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
data/06_fund/train.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
data/07_cf/train.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
data/08_calc/train.jsonl ADDED
The diff for this file is too large to render. See raw diff