EnvLoop/financial-decision-bench

EnvLoop Financial Decision Bench

A multilingual, source-grounded collection of financial decision tasks with explicit evidence, constraints, choice labels and provenance. This is the public v1.0.0 view of the frozen matrix-v1 research corpus, released on 2026-10-21.

Start with default. It contains 159,000 training rows and 15,600 validation rows, representing 53,000 / 5,200 base cases in English, Chinese and Japanese. The separate crypto-nc configuration contains 1,500 / 300 rows and is CC BY-NC-SA 4.0, noncommercial only. Across both non-overlapping configurations there are 176,400 language rows / 58,800 base cases. Language-specific configurations are views of default, not additional data.

Load the data

from datasets import load_dataset

train = load_dataset(
    "EnvLoop/financial-decision-bench", "default",
    split="train", revision="v1.0.0", streaming=True,
)
example = next(iter(train))
prompt_messages = example["messages"]  # system + user only
answer = example["answer"]             # original option ID, e.g. "C"

# For supervised training only:
training_messages = prompt_messages + [{"role": "assistant", "content": answer}]

Use "en", "zh" or "ja" instead of "default" to select one language. split="validation" loads the original DEV split. Data use standard Parquet, with no remote dataset loading code required. The release was checked with datasets==5.1.0, pyarrow==26.0.0, and huggingface_hub==2.2.0.

Configurations and splits

Configuration	Train rows	Validation rows	Scope
default	159,000	15,600	11 abilities, three languages; source-specific terms apply
en, zh, ja	53,000 each	5,200 each	Views of default; do not add these counts again
crypto-nc	1,500	300	Binance Vision funding-derived tasks; noncommercial, share-alike

The original local matrix has 198,000 language rows, including a separate 16,500-row final TEST. No final TEST questions or labels are included in this repository. This public release also omits all 5,100 Coinbase-derived TRAIN/DEV rows. The original full archive, legacy releases, raw provider archives and prepared token SQLite are not published here because they include held-out answers, restricted source material or incompatible licenses.

The publication filter uses source provenance and upstream conditions, never model correctness. Names of Coinbase entities/exchanges appearing in SEC/CFTC material are retained: those facts originate from the government source, not Coinbase market-data feeds. A World Bank indicator with unverified indicator-level permission (GC.BAL.CASH.GD.ZS) is excluded by policy; no selected matrix TRAIN/DEV row depended on it.

Capabilities

The 11 ability IDs are credit, fixed_income, portfolio, fx_liquidity, options, energy, declared_condition_application, information_timing_scope, measurement_interpretation, official_decision_selection, and unit_and_parameter_mapping.

After publication filtering, counts are no longer 5,000 TRAIN rows in every language/ability cell. Actual counts per language are below; the three language views are balanced with each other.

Ability	Train per language	Validation per language
credit	5,000	500
declared_condition_application	5,000	500
energy	5,000	500
fixed_income	5,000	500
fx_liquidity	5,000	500
information_timing_scope	5,000	500
measurement_interpretation	4,250	400
official_decision_selection	5,000	500
options	5,000	500
portfolio	5,000	500
unit_and_parameter_mapping	3,750	300

Fields and training contract

• id is the original row ID; base_case_id, translation_group_id and source_group_ids support grouping. Keep all translations and linked sources in the same split.
• messages contains the original shared prompt contract: one system message and one user message, with evidence, conditions, question and original option order. It excludes answer metadata.
• answer is one original uppercase choice ID. options is an ordered list of {id, text}. Use the original IDs rather than renumbering or sorting options by predicted probability.
• state_json and question_json preserve the model-visible JSON as strings so heterogeneous financial structures have a stable table schema. Decode them with json.loads if needed.
• language, ability, construction_type, difficulty and source_dataset describe the task. Difficulty is a construction label, not an empirically guaranteed error rate.
• reference_date, publication_time_json, and dates in the evidence/provenance have different meanings. Missing values are retained; reference dates are not proof of point-in-time public availability.
• source_urls and provenance_json retain original URLs, locators and source hashes. Project-local file paths were removed from provenance. source_terms identifies the applicable source conditions.
• prompt_sha256, original_input_sha256 and original_target_sha256 bind the export to the frozen input/target records. The target hash is a binding, not a copy of hidden TEST labels.

Only append the answer for supervised TRAIN preparation. For inference, send messages without answer or other target metadata. The original Qwen3.8 recipe supervises only the final option-label token, masks prompt loss and adds no EOS loss. Tokenizer-specific checks are still required; this repository does not distribute model weights or claim a trained model result.

Validation baselines

The table below re-scores previously frozen predictions on exactly the public default validation IDs. No new model requests were made for this release. Micro accuracy counts invalid/missing answers as incorrect; the macro column weights the 33 language/ability cells equally.

Service-reported model	Correct / planned	Micro accuracy	Cell-macro accuracy	Valid coverage
gpt-6-astra	15,570 / 15,600	99.81%	99.82%	100.00%
jev-1.14.0	9,064 / 15,600	58.10%	59.88%	99.85%

See baseline metadata for per-cell figures, the distinct crypto-nc results, and frozen prediction hashes. Service model labels do not independently verify underlying weights. These are development results, with prior development reuse; they are not a final held-out comparison, proof of trading profit, or evidence that a fine-tuned model outperforms either baseline.

Sources and permissions

Default uses structured official facts and scoped text from SEC, CFTC, EIA, Federal Register/GovInfo and OFAC; ECB reference rates; New York Fed reference rates; and individually reviewed World Bank WDI indicators. Authored scenarios may combine those real observations with explicitly stated hypothetical fees, thresholds, portfolios, stress paths or business constraints. Such assumptions and derived labels are not claimed as observed issuer or trader decisions.

Read LICENSE and SOURCE_NOTICES.md. Original benchmark authorship is available under CC BY 4.0, while upstream source permissions and required notices remain applicable. The separate crypto-nc configuration is subject to CC BY-NC-SA 4.0, including its derived task/label material, and must not be silently mixed into a commercial training corpus. Coinbase data and dependent tasks require separate permission and are absent.

New York Fed reference-rate observations are subject to the Terms of Use posted at newyorkfed.org. The New York Fed is not responsible for publication of these data by EnvLoop, does not sanction or endorse any particular republication, and has no liability for your use. EnvLoop is not affiliated with the New York Fed. The New York Fed does not sanction, endorse, or recommend any products or services offered by EnvLoop. Additional upstream notices are retained in SOURCE_NOTICES.md.

No source institution endorses this dataset. Official values, reporting scopes, units and dates remain distinguishable from author calculations and translations. No private vendor feed, issuer narrative archive or government seal is licensed by the benchmark's authored-content license.

Limitations

• The same base economic case is translated into three languages. The source-language metadata are English; this is not equal native-source coverage of three countries.
• Base cases can share filings, entities, source observations and construction templates. Counts are not counts of independent market events; zero cross-split IDs/group IDs is not a guarantee of all forms of statistical independence or pretraining decontamination.
• Some tasks are exact calculations or condition application with high frontier-model accuracy. Hard-design tags do not establish a difficult frontier benchmark.
• Historical timestamps, source revisions, convenience selection and explicit synthetic assumptions limit point-in-time financial claims. OFAC rows describe historical record actions, not current sanctions-screening advice.
• This public validation split is already exposed and was used in development. Future claims should use a separately protected, newly frozen final evaluation.

Files and verification

The 45 Parquet shards are described in metadata/release.json. metadata/distribution.json, metadata/schema.json, and metadata/source-policy.json document row counts, fields and the exact publication filter. SHA256SUMS covers the published release files. The export preserves input values, option order and labels while removing local provenance paths and naming DEV as validation.

中文使用说明

默认使用 default 配置：159,000 条训练数据、15,600 条验证数据，覆盖中英日和 11 种能力。en、zh、ja 是同一批题的语言视图。messages 只含模型输入，answer 是单个选项 ID；仅在训练时把答案追加为 assistant 输出。

另有 crypto-nc 配置，共 1,800 条，只适用于其非商业、相同方式共享许可。Coinbase 来源题目已剔除，最终 TEST 未公开。公开版共 176,400 条语言样本，不是原始全部 198,000 条，也不再保证每个能力×语言训练格恰好 5,000 条。真实来源、模拟条件、来源许可和验证集重复使用的边界见上文与 SOURCE_NOTICES.md。