Modelpublic

spkc83/retail-bank-servicing-agent-9b-peft-v12-honesty

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes28downloads

Model Card

Retail Bank Agent 9B

This is a research checkpoint for a synthetic retail-bank customer-service demonstration. It is a bf16-lora LoRA adapter (rank 32, alpha 64) trained on top of spkc83/retail-bank-servicing-agent-9b at revision 1d56824995aa1adecfe20f62ca42fb1c0c443817; the base has approximately 8.8 billion parameters and uses its native tagged JSON tool-call format.

Training

—Dataset: spkc83/retail-bank-servicing-alignment-sft at 8494c94f9da4ada0a26de988781b88cc2ec58c53
—Training records: 3959
—Validation records: 447
—Tool manifest: nine synthetic retail-banking tools
—Assistant-only target masking: tool-call and final-assistant spans
—Maximum sequence length: 2048
—Optimizer steps: 2000
—LoRA rank/alpha: 32/64
—Source commit: 652b2961570537fb71a5abe6bfd8f1a247e1d981
—Chat-template SHA-256: 6727ca16a39df05c41af54eb651aa618b50a29967ad3951a31b90c4e385573fc

The repository root is the trained BF16 LoRA adapter only: there are no merged weights and no config.json. Load it with PeftModel.from_pretrained(base, repo, revision=...) on top of the base model at the pinned revision. The same adapter is duplicated under adapter/.

Intended use and limitations

The model is intended only for the linked synthetic banking POC. It must be given the published tool schemas and tool results. It has no access to real banking systems, is not financial advice, and may make incorrect tool choices or unsupported claims. Evaluate tool-call syntax, arguments, backend execution, grounded final responses, OOD behavior, and multi-turn behavior before relying on a revision.