Etherlabs/finance-ops-triage-v0.1

Finance Ops Triage v0.1 dataset

The original small, illustrative dataset prepared for Ugo Chukwu's first Unsloth fine-tuning and deployment exercise. The examples were provided during a guided ChatGPT experiment; they are not collected operational transaction records or an independently validated finance policy.

Code and experiment report · Model archive

Structure

Each JSONL row has messages containing system, user, and assistant entries. The assistant content is a JSON string with exactly category, severity, and next_action.

Split	Rows	Categories	Severity counts
train	50	10 each	high 23, medium 17, low 10
validation	15	3 each	high 9, medium 3, low 3

Categories: settlement_mismatch, duplicate_transaction, missing_transaction, refund_issue, other. Severities: low, medium, high.

The file named eval served as validation during training. Five manual before/after cases are validation rows 1, 4, 7, 10, and 13. There is no untouched test split. There are no exact train-validation incident-text duplicates, but many examples are semantically similar. The archived files match the copies uploaded to Unsloth byte for byte; hashes are included.

Intended use and limitations

Use as a transparent educational record or a tiny development dataset. It is too small and repetitive to establish operational reliability. Severity rules are incompletely specified and inconsistent in places. other is dominated by reporting requests. Category boundaries, hard negatives, and phrasing coverage are limited. Targets contain simplified actions that are not authorization to modify real records.

Do not interpret the historical 40% → 60% composite rubric as broad classification accuracy: category correctness was 3/5 both before and after fine-tuning, and severity correctness declined. Preserve this original dataset when constructing a better v0.2 with explicit rules and protected test data.

The examples use illustrative amounts, provider names, and transaction references. No customer identity or operational export is included. Original data licensing has not yet been selected; public availability alone does not establish an unrestricted reuse license.