Modelpublic

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

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes9downloads

Model Card

Retail Bank Servicing Agent 9B — PEFT Adapter

This repository contains the Granite 9B LoRA adapter only. Inference composes it with the exact base checkpoint below using BF16 base weights; it does not contain merged 9B weights.

—Final immutable adapter-bundle revision: b4269445ce7b2b943d2d9531102166bf8840a074
—Base model: spkc83/retail-bank-servicing-agent-9b
—Base revision: 1d56824995aa1adecfe20f62ca42fb1c0c443817
—SFT dataset: spkc83/retail-bank-servicing-alignment-sft
—Dataset revision: 40a0b68b9f746131ffff32a83e077fd7e4a344d1
—Dataset manifest SHA256: 6c0f8c6bb8d0e1d646c4d43226aef9fba369a891c9ea7a3bb58764e07d6e9bbe
—Source commit: 75b56ffff45e75ffbee11c0e0552dc35ae124d21
—Training job: 6a7f79531f5885ae605b96cc
—Optimizer steps: 750
—Training loss: 0.13014758084217706
—Validation loss: 0.32008039951324463
—Validation token accuracy: 0.9624034788297571
—Adapter SHA256: 043b22c5203b222522de6e47ace7da0a32a4119a053ec76269a3d8159e4ebd81
—PEFT composition: LoRA rank 32, alpha 64, dropout 0.05, BF16 base weights

Load the pinned base revision first, then attach this adapter revision with PEFT. The finalizer's JSON report distinguishes the immutable adapter-bundle commit from the metadata-only repository-head commit.

This model is experimental, trained on synthetic retail-banking data, and has no connection to a real bank.