andy279/finjev-financial-decisions

FinJev Financial Decision Development Set — v0.1.0

Private initial release by andy279, with project code tracked at https://github.com/nhannguyen2709/finjev.

60 unique questions, 90 representation rows. All labels are provisional, machine reviewed, and not independently expert-adjudicated. This is development data, not a protected test or representative estimate of financial model performance.

Configurations

Configuration	Rows	Purpose
pilot30	30	Initial easier pilot, ten cases per task
hard30_html	30	Harder challenge, ten cases per task, compact HTML evidence
hard30_structured	30	The same hard questions and labels, with structured table arrays

The two hard configurations have identical case_id values. They are paired representations, never independent examples or separate train/test splits. The issuers overlap the easier pilot. No training/validation/test split is claimed.

Three decision tasks

1. Evidence selection: choose A, B, C or NONE. For hard cases each candidate is a bundle; judge it using only its assigned evidence. The bundle must establish every material part of the claim.
2. Claim verification: SUPPORTED, CONTRADICTED or INSUFFICIENT, using the supplied company, period, units, measure and scope. Missing evidence is not contradiction. Explicit arithmetic and unit conversion are allowed.
3. Reporting-period comparability: COMPARABLE, NOT_COMPARABLE or INSUFFICIENT under the rule stated in the question. A known violation takes precedence over missing information. This does not predict financial importance or stock returns.

This release tests decisions with supplied evidence. It does not test corpus retrieval or expert agent trajectories. NONE/INSUFFICIENT are answer categories, distinct from low confidence or abstention.

Schema and loading

question and label are the unchanged frozen strings. criteria_json contains answer definitions; state_json contains evidence, metadata, citation locations and candidate bundles. reference_json contains rationale, supporting quotes, annotation provenance and review status. These three fields are JSON strings to support variable table/criteria schemas without lossy Arrow coercion. Parse them with json.loads.

import json
from datasets import load_dataset
rows = load_dataset("andy279/finjev-financial-decisions", "hard30_structured",
                    split="development", revision="v0.1.0", token=True)
example = rows[0]
state = json.loads(example["state_json"])
criteria = json.loads(example["criteria_json"])
# Supply state, question and criteria to the model. Never supply label/reference_json.

Use case_id to group duplicate representations; example_id is only unique within a configuration. original_input_sha256 binds to the original local input, before portable path substitution. The release manifest hashes the actual exported files. provenance/ includes original freeze manifests and source-file hashes. Local paths are mapped reversibly to archive://financial/, project:// and user-home:// aliases; these are identifiers, not downloadable URLs.

Sources and annotation

Evidence comprises excerpts and tables from deliberately selected SEC company filings, with company/CIK, accession, document, filing date, reporting period and HTML-body XPath where available. Filing dates are catalog dates; public acceptance timestamps were not independently verified. Source hashes identify the archived derivatives, not authenticated SEC originals. See provenance manifests for exact coverage and hashes.

Claude proposed questions, difficult alternatives and labels and performed review. Deterministic checks validated cited local source text/table structure and hashes; a Codex-agent semantic audit refined ambiguous constructions. Neither model-family consensus nor matching a source establishes expert gold labels. Claude's author/reviewer role can advantage it during evaluation.

Questions and labels were frozen before evaluations. The structured-table transformation preserved source text, choices and labels, replaced HTML with row arrays and merged ranges, and retained the original flattened text. It reduced input length as well as changing format.

Recorded results (agreement with provisional labels)

Model	Pilot30	Hard HTML	Hard structured
jev-1.13.0	30/30	10/30	12/30
claude-opus-5-5	30/30	30/30	30/30
gpt-6-astra	30/30	29/30	30/30

CLI models used high reasoning effort. Immutable CLI weight revisions are unavailable. Jev returns native API probabilities; Claude/Codex probabilities are self-reported. One Jev structured response (P022) summed to 0.99 and was rejected by the unchanged validator (tolerance 1e-5), retained as a failure and counted incorrect. No retry or normalization was used. Its raw answer also disagreed with the reference. Evaluation files preserve valid outputs and failure records, with no raw CLI transcripts/reasoning logs. Probability metrics exclude invalid responses; the paired comparison reports the common-valid subset. Source snapshots were not selected based on evaluation failures.

These are single runs, with correlated issuer/source groups and no matched repeated HTML control. They cannot establish a definitive model ranking, financial calibration or a causal formatting effect. Accuracy/F1, Brier sum, log loss, calibration summaries and risk/coverage appear in evaluation files. Latency includes backend/CLI overhead and is not matched inference compute.

Known limitations and unresolved issues

• Hard P012's NOT_COMPARABLE versus INSUFFICIENT distinction requires independent adjudication. The original frozen label is retained in both representations.
• The hard comparability subset has five COMPARABLE and five NOT_COMPARABLE labels, with no INSUFFICIENT examples.
• No issuer-held-out or forward-time performance claim. All examples are deliberately selected; the hard set is a challenge set, not representative sampling.
• Public filings may have appeared in model pretraining. No claim of contamination-free evaluation.
• Existing section labels, similarity scores and machine-proposed labels are not ground truth.
• Source redistribution/training rights have not been fully reviewed. This first release remains private; see RIGHTS.md.

Before scaling, independently adjudicate labels, group source/issuer/near-duplicates across splits, build forward-time and issuer holdouts, separate representative and challenge evaluations, and protect future test labels from training. No new evaluation or training is part of this upload.