spkc83/retail-bank-agent-sft

Retail Bank Agent Tool-Use SFT

This dataset contains 9,000 deterministic, fictional retail-banking conversations for supervised fine-tuning of a conversational tool-using model.

• —Dataset: https://huggingface.co/datasets/spkc83/retail-bank-agent-sft
• —Training revision: 183e7e1ed1aba9c3d7155e7b83b64dc854935055
• —Source: https://github.com/spkc83/retail-bank-servicing
• —Model: https://huggingface.co/spkc83/retail-bank-agent-9b
• —Public POC: https://huggingface.co/spaces/spkc83/retail-bank-servicing-poc

Splits

• —Train: 6,304
• —Validation: 1,349
• —Frozen test: 1,347
• —Corpus fingerprint: 2bb7a400ed2556b15c7e5eb6147668041b5deef8ae4f037f9e2e52295ff29ab5
• —Split seed: 711

Coverage

The corpus covers all nine public synthetic-bank tools, successful and failed tool results, clarification, general banking FAQ, hard-negative private-field requests, out-of-domain refusal, multi-turn context, and ordered multi-tool calls.

Every tool-bearing record was replayed against isolated deterministic synthetic state before inclusion. Assistant tool-call and final-response tokens are trainable; system, user, and tool-result tokens are context only. The data validator rejects semantically empty final responses and asserts path-specific content for clarification, FAQ, OOD, and hard-negative rows.

Source and privacy policy

All included rows are self-authored synthetic data under MIT. Banking77 and CLINC are classifier/evaluation-only and contribute no generative SFT rows. Bitext remains quarantined and contributes no rows.

The dataset contains no real customers, credentials, accounts, or financial events. It is for research demonstrations, not production banking.