CentificAIResearch/tiered-finance-eval

Tiered Finance Eval

Twenty agentic finance tasks, each with the reference files an analyst would actually be handed, a curated gold deliverable, and a tiered, gated rubric that scores a submission against that gold.

Evaluation results for these tasks are published in the companion Space: CentificAIResearch/Tiered-Finance-Eval. This dataset holds the tasks only: no model outputs and no scores.

Canary string. TIERED-FINANCE-EVAL:d9e2f4a1-7c3b-4e86-9a05-2f1b8c6d40e7 If a model reproduces this string, this dataset reached its training data. Please do not strip it, and please exclude this repository from pretraining crawls.

What a task looks like

A task is a realistic piece of finance work — build a credit case for a hospital revenue bond, run a bank safety-and-soundness exam, reconcile an SBA 7(a) loan payoff, stress-test a portfolio. The model gets a prompt and a set of reference_files (real filings, regulator PDFs, FRED series, agency spreadsheets) and must produce a deliverable in the format the prompt asks for — most often a Word memo or an Excel model, sometimes a PDF, deck or markdown report.

Tasks	20
Occupations	7
Sector	Finance and Insurance
Reference files	40 — pdf 14, csv 10, html 10, xlsx 5, txt 1
Gold deliverable files	30 — docx 11, xlsx 10, pdf 4, txt 2, md 1, pptx 1, csv 1
Rubric items	1,484 across the 20 tasks (39–100 per task)

Occupations: Financial and Investment Analyst (2) · Personal Financial Advisors (2) · Financial Examiners (3) · Insurance Underwriters (3) · Credit Analysts (2) · Bookkeeping, Accounting, and Auditing Clerks (4) · Financial Risk Specialists (4).

Columns

Column	Type	Notes
task_id	string	e.g. financial_examiners_7320
sector, occupation	string	O*NET-style role labels
prompt	string	The full task instruction given to the model
reference_files	list[string]	Paths in this repo, ready to pass to hf_hub_download
reference_file_urls	list[string]	Direct https://huggingface.co/.../resolve/main/... download URLs
reference_file_hf_uris	list[string]	hf://datasets/... URIs for huggingface_hub
reference_file_source_urls	list[string]	Where each file was originally obtained (SEC, FDIC, SBA, FRED, CMS, …)
deliverable_files	list[string]	The gold deliverable, same path convention
deliverable_file_urls, deliverable_file_hf_uris	list[string]	As above. Gold files are authored, so they have no source URL
rubric_pretty	string	The rubric as readable text
rubric	list[object]	The rubric as structured items — see below

Everything is plain JSON, so you do not need a dataframe library to read it:

import json
tasks = json.load(open("data/tasks.json"))        # a list of 20 task objects
print(tasks[0]["rubric"][0]["criterion"])

It also loads as a normal dataset. There is one split and it is named test: this is an evaluation set, not training data.

from datasets import load_dataset

ds = load_dataset("CentificAIResearch/tiered-finance-eval", split="test")
row = ds[0]
print(row["task_id"], len(row["rubric"]), "rubric items")

To fetch a task's input files, pass the paths straight to hf_hub_download:

from huggingface_hub import hf_hub_download

for path in row["reference_files"]:
    local = hf_hub_download("CentificAIResearch/tiered-finance-eval",
                            filename=path, repo_type="dataset")

File layout

data/tasks.json                                  all 20 tasks, one JSON array
tasks/<task_id>/task.json                        the same record, beside its files
tasks/<task_id>/reference_files/<hash>/<file>
tasks/<task_id>/deliverable_files/<hash>/<file>

Files are nested per task, so each task folder is self-describing and no two tasks can share or overwrite a path. file_manifest.json lists every published file with its size and SHA-256.

The rubric

Each item is scoped to one of three nested tiers — T1 Domain ⊃ T2 Occupation ⊃ T3 Task — so a submission is scored on general finance competence, occupational competence and task-specific execution separately, rather than against one flat checklist.

Every item has exactly these four fields:

{
  "criterion": "Do all material quantitative claims fall within a defensible tolerance band …",
  "tier": "T1",
  "critical": true,
  "weight_range": [0, 5]
}

• tier — T1 and T2 items score on a 0–5 scale, T3 on 0–10. The 1,484 items split 302 / 441 / 741.
• critical — a gating item. Failing one caps the score regardless of what else was done well; this is what separates "produced a plausible-looking document" from "produced a usable one".
• weight_range — [0, n] for a positive criterion; a range whose lower bound is negative marks a guardrail: a penalty for a specific serious error (fabricating a figure, citing a non-existent source, presenting stale data as current) rather than for the mere absence of something good. 129 of the 1,484 items are guardrails. Read polarity from weight_range, never from the sign of a reported award.

Scope and intended use

• These tasks are demanding by design. They were chosen to be discriminating rather than typical, so scores on this set run lower than on everyday finance work and are not an unbiased estimate of general performance. Use the set to compare systems against each other, not to characterise a system's competence in absolute terms.
• Rubrics are model-generated and programmatically validated, assembled from structured per-tier criterion banks. They have not been individually reviewed by a domain expert, and some items are anchored to the particular structure of the gold rather than to the underlying requirement.
• Gold deliverables are curated reference answers, not formal ground truth. More than one competent analyst answer usually exists for a task; treat the gold as a high-quality reference point rather than the only correct output.

Provenance and licensing

The prompts, gold deliverables and rubrics are original work, released under CC-BY-4.0.

The reference files are redistributed copies of publicly available source documents — US government and regulator publications (SEC, FDIC, OCC, SBA, CMS, EIA), Federal Reserve (FRED) data series, and public audit reports. Each file's origin is recorded in reference_file_source_urls. They are included so the tasks stay runnable as their original links rot; rights in those documents remain with their issuers. If you are the rights holder of a redistributed file and would like it removed, please open a discussion on this repository.

Changelog

• 2026-10-06 — Rubrics moved from four tiers to three: the former T2 Persona and T3 Occupation tiers are merged into T2 Occupation, and the former T4 Task tier is now T3. Every criterion, critical flag and weight_range is unchanged, as are the 20 tasks, prompts, reference files and gold deliverables. The persona column is removed.
• 2026-09-18 — Initial release.

Citation

@misc{tiered_finance_eval_2026,
  title  = {Tiered Finance Eval: agentic finance tasks with gold deliverables and tiered gated rubrics},
  author = {Centific AI Research},
  year   = {2026},
  url    = {https://huggingface.co/datasets/CentificAIResearch/tiered-finance-eval}
}