Blame for README.md

---
license: cc-by-4.0
language:
- en
- zh
pretty_name: FinForesight
tags:
- finance
- benchmark
- deep-research
- agent-evaluation
size_categories:
- 1K<n<10K
configs:
- config_name: val
  data_files:
  - split: train
    path: data/01_val/train.jsonl
  default: true
- config_name: rank_f
  data_files:
  - split: train
    path: data/02_rank_f/train.jsonl
- config_name: rank_t
  data_files:
  - split: train
    path: data/03_rank_t/train.jsonl
- config_name: rank_h
  data_files:
  - split: train
    path: data/04_rank_h/train.jsonl
- config_name: evt
  data_files:
  - split: train
    path: data/05_evt/train.jsonl
- config_name: fund
  data_files:
  - split: train
    path: data/06_fund/train.jsonl
- config_name: cf
  data_files:
  - split: train
    path: data/07_cf/train.jsonl
- config_name: calc
  data_files:
  - split: train
    path: data/08_calc/train.jsonl
---

# FinForesight

Evaluation dataset for financial Deep Research Agents (**data only**).

Code, builders, prompt templates, and evaluation scripts live in a separate repository. This Hub repo publishes standardized ready seeds only, loadable by subset.

Chinese documentation: [`README_cn.md`](./README_cn.md).

## Links

- **Code**: https://github.com/OpenFinMind/FinForesight
- **Dataset**: https://huggingface.co/datasets/OpenFinMind/FinForesight

Task configs: `val` / `rank_*` / `evt` / `fund` / `cf` / `calc`
(Valuation / Return Ranking / Event Direction / Fundamentals / Counterfactual / Calculation).

## Quick Start

```python
from datasets import load_dataset

# Load by task subset (recommended)
val = load_dataset("OpenFinMind/FinForesight", "val")
fund = load_dataset("OpenFinMind/FinForesight", "fund")
cf = load_dataset("OpenFinMind/FinForesight", "cf")

print(val["train"][0]["task_id"])
print(val["train"][0]["prompt"][:200])

# Filter by time band
t2 = val["train"].filter(lambda x: x["time_band"] == "T2")
```

Available configs: `val` / `rank_f` / `rank_t` / `rank_h` / `evt` / `fund` / `cf` / `calc`  
Default config: `val`.

### EVT / FUND perturbations

`evt` and `fund` include perturbation items. Each is a rewrite of a base item in the same category, asking the same question with identical `ground_truth`, used to measure answer stability under rewrite.

- `variant` is `perturb_layout` (layout reshuffle) / `perturb_paraphrase` (instruction paraphrase) /
  `perturb_numeric` (disclosure numbers scaled by the same factor ±1%); other `variant` values are base items.
- `metadata.perturbation.base_task_id` points to the base item; `metadata.perturbation.perturbation_type`
  gives the perturbation type.
- To evaluate base items only, filter on `variant`:

```python
evt = load_dataset("OpenFinMind/FinForesight", "evt")
base = evt["train"].filter(lambda x: not x["variant"].startswith("perturb_"))
```

## Scale (v0.9.0)

| config | task | ready count |
|--------|------|------------:|
| `val` | Valuation (VAL) | 244 |
| `rank_f` | Return Ranking F (RANK-F) | 112 |
| `rank_t` | Return Ranking T (RANK-T) | 161 |
| `rank_h` | Return Ranking H (RANK-H) | 110 |
| `evt` | Event Direction (EVT) | 228 |
| `fund` | Fundamentals (FUND) | 384 |
| `cf` | Counterfactual (CF) | 30 |
| `calc` | Calculation (CALC) | 30 |
| **total** | | **1299** |

### Experiment identity (paper primary)

- `backbone_model = gpt-4.1`
- `backbone_training_cutoff = 2024-06-30` (OpenAI states “June 2024” only; this release uses month-end conservatively)
- `guard_days = 30` → T1 outcome ≤ `2024-05-31`, T2 origin ≥ `2024-07-30`
- `experiment_as_of = 2026-08-17`

**Incompatibility note:** the legacy Hub pin (339 items / `model_training_cutoff=2024-06-01`) is a legacy identity and must not be mixed with v0.9 `paper_band` / C1 tables.

## Fields

One evaluation item per row (JSON object):

| field | description |
|------|-------------|
| `task_id` | Globally unique ID |
| `category` | VAL / RANK / EVT / FUND / CF / CALC |
| `variant` | Sub-variant (e.g. F/T/H, earnings) |
| `cutoff_date` | Cutoff used for temporal banding |
| `time_band` | T1 / T2 / T3 (legacy; use `paper_band` for papers) |
| `forecast_origin` | Forecast origin time |
| `outcome_available_at` | Outcome availability date; use with `outcome_available_at_source` for observed vs estimated |
| `paper_band` | GPT-4.1 paper banding: T1 / T2 / T3 / CF / CALC |
| `status` | ready |
| `seed` | Structured input |
| `prompt` | Fully rendered prompt |
| `expected_output` | Output schema description |
| `ground_truth` | Evaluation label |
| `metadata` | Metadata |

## Temporal banding

This dataset labels legacy `time_band` with a reproducible rule (based on `cutoff_date` only):

- `T1`: `cutoff_date <= model_training_cutoff` (possible pretraining leakage)
- `T2`: `model_training_cutoff < cutoff_date < reference_current_date`
- `T3`: `cutoff_date >= reference_current_date` (future-forecast stress)

Default legacy parameters (aligned with paper cutoff month, different rule):

- `model_training_cutoff = 2024-06-30`
- `reference_current_date = 2026-08-17`

For paper experiments, use sidecar `paper_band` (see code repo `docs/schema.md` / `docs/c1_pairing_contract.md`); do not use legacy `time_band` for C1.

## Repo boundary

This repository **only** contains:

- `README.md` (English dataset card)
- `README_cn.md` (Chinese dataset card)
- `data/<config>/train.jsonl` (ready seeds)

It does **not** include code, scripts, prompts, raw build CSVs, or run outputs.

## License

[CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)