README.md

metadata

license: other
license_name: mixed-rights
license_link: https://github.com/wjh19990923/fred-sembench/blob/v1.0.0-paper/DATA_LICENSE.md
language:
  - en
pretty_name: FRED-SemBench
size_categories:
  - n<1K
task_categories:
  - question-answering
tags:
  - finance
  - macroeconomics
  - benchmark
  - agent-evaluation
  - semantic-reliability
configs:
  - config_name: questions
    data_files: questions.jsonl
  - config_name: facts
    data_files: facts.jsonl
  - config_name: llm_only_responses
    data_files: llm_only_responses.jsonl
  - config_name: fred_web_responses
    data_files: fred_web_responses.jsonl
  - config_name: iris_responses
    data_files: iris_responses.jsonl
  - config_name: llm_only_scores
    data_files: llm_only_scores.jsonl
  - config_name: fred_web_scores
    data_files: fred_web_scores.jsonl
  - config_name: iris_scores
    data_files: iris_scores.jsonl
  - config_name: llm_only_tool_traces
    data_files: llm_only_tool_traces.jsonl
  - config_name: fred_web_tool_traces
    data_files: fred_web_tool_traces.jsonl
  - config_name: iris_tool_traces
    data_files: iris_tool_traces.jsonl
  - config_name: iris_coverage_scores
    data_files: iris_coverage_scores.jsonl
  - config_name: iris_supported_questions
    data_files: iris_supported_questions.jsonl
  - config_name: source_provenance
    data_files: source_provenance.jsonl

FRED-SemBench

Code & Evaluation (GitHub) | Paper (ACL Anthology) | Paper-aligned Release

FRED-SemBench is a 200-question benchmark candidate for evaluating whether LLM agents retrieve macroeconomic answers with the intended concept, series, transformation, unit, observation period, and data-vintage semantics.

This dataset accompanies the FinNLP 2026 paper “FRED-SemBench: Evaluating Semantic Reliability in LLM Access to Macroeconomic Data” by Wilson Wang, Chandler Han, and Peter Zhang (Kairos-AI).

Status and scope

• 50 independently selected macroeconomic concepts
• 100 verified derived fact records
• 200 machine-generated benchmark questions
• 48-question three-condition diagnostic evaluation
• 22-question pre-registered Iris-supported primary slice
• Full automated validation completed
• Independent domain adjudication remains future work

The dataset should therefore be treated as a paper-aligned benchmark candidate, not as a fully adjudicated production benchmark.

Quick start

from datasets import load_dataset

questions = load_dataset(
    "wangjinh/FRED-SemBench",
    "questions",
    split="train",
)

Configurations

The Hub viewer exposes separate configurations for questions, facts, each condition's recorded responses, deterministic scores, credential-free tool traces, and Iris coverage artifacts. Additional CSV/JSON metadata files are available from Files and versions.

The evaluated conditions are:

1. llm_only: no external data tools;
2. fred_web: live Web search and a read-only FRED API wrapper;
3. iris: the Kairos-AI Iris semantic layer and CLI, with direct FRED, Web, and shell access disabled.

Paper results

Scope	LLM-only	FRED+Web	Iris
Iris-supported slice	4/22	13/22	17/22
48-question diagnostic slice	4/48	28/48	25/48

The complete validation and reproduction code is available at:

https://github.com/wjh19990923/fred-sembench

The immutable paper-aligned artifact is available at:

https://github.com/wjh19990923/fred-sembench/releases/tag/v1.0.0-paper

Data fields

questions.jsonl contains the natural-language question, concept and task identifiers, explicit semantic gold contract, source attribution, precision, scope expectations, confounders, and annotation status. See DATA_DICTIONARY.md for the field-level description.

The response files contain the recorded final answer and operational metadata. The score files contain deterministic classifications and component checks. The trace files contain observable, allowlisted tool-call metadata and hashes, not tool result bodies or hidden model reasoning.

Privacy and commercial-system boundary

Per-run Codex thread identifiers were removed before publication. The release does not contain API keys, authentication files, cookies, local runtime homes, or hidden chain-of-thought.

Iris is a commercial Kairos-AI system. This release includes its paper prompts, recorded responses, deterministic scores, coverage decisions, and credential-free traces. It does not include the proprietary production implementation or an entitlement to the commercial service.

FRED source-data notice

Raw FRED CSV snapshots are intentionally not redistributed. The release retains series identifiers, source URLs, retrieval dates, derived evidence, and SHA-256 hashes for research auditability.

This product uses the FRED® API but is not endorsed or certified by the Federal Reserve Bank of St. Louis. Individual series may carry source-specific copyright and redistribution conditions. Users must comply with the current FRED terms and each original provider's terms:

• https://fred.stlouisfed.org/legal/
• https://fred.stlouisfed.org/docs/api/terms_of_use.html

See DATA_LICENSE.md for the mixed-rights notice.

Citation

Please cite the accompanying FinNLP 2026 paper, available on ACL Anthology:

https://aclanthology.org/2026.finnlp-1.19/

Citation files and evaluation code are maintained in the GitHub repository.

Contact

Peter Zhang, corresponding author — founders@kairos-ai.io