arpjw/Nemotron-Mini-4B-Banking

Nemotron Mini 4B Banking

This is an independently trained adapter only for NVIDIA's Nemotron-Mini-4B-Instruct, pinned to base revision 791833e92ebddb0bc2c1007f6d2b6764f886a2ae. It classifies short banking-support requests into BANKING77's fixed set of 77 intents. Arya Somu trained the adapter; NVIDIA did not make or endorse this release. The package contains no base weights, tokenizer, official-test queries, or per-query predictions.

Measured result

The selected rank-16 LoRA adapter trained for one epoch on 2,308 examples from the publisher's training partition. Selection used a separate 1,993-example development split. Under the same taxonomy-constrained greedy decoder, the untouched base reached 0.41337 macro-F1 and 1,310/3,080 accuracy on the official test. The adapter reached 0.84443 macro-F1 and 2,614/3,080 accuracy. The paired macro-F1 gain was 0.43105 with a 95% normalized-query-group bootstrap interval of [0.41517, 0.44919]. This is one seed on a historical, in-taxonomy benchmark. It is not evidence of general finance skill or production banking safety.

The source and protocol describe the fixed prompt, complete label taxonomy, constrained decoder, preprocessing, training, and evaluation. The result report documents the gate, official comparison, overlap sensitivity, limitations, and artifact hashes. The article explains the experiment for general readers.

Use and limits

Load the NVIDIA base model and tokenizer at the exact revision above, then attach adapter_model.safetensors using adapter_config.json with PEFT 0.15.2. The adapter config names the base repository but does not itself pin its revision, so pin the base explicitly. For the reported result, use the prompt, 77-label taxonomy, decoding, and evaluation code in the linked source repository. A generic text-generation call is not the reported classifier evaluation.

The model must choose one of 77 labels. It was not tested for rejection of other intents, calibrated confidence, unfamiliar banks, financial advice, or customer-facing safety. A TF-IDF logistic classifier trained on the same examples reached 0.82852 development macro-F1, near the adapter's 0.82978. The public benchmark may have appeared in base-model pretraining; that could not be ruled out. Five official rows were flagged for capped-train/development overlap or within-test label conflict; the all-row result is primary and the sensitivity check is in the report.

Data and licensing

BANKING77 is from PolyAI and the authors of Efficient Intent Detection with Dual Sentence Encoders, licensed CC BY 4.0. The training subset was capped, deduplicated by normalized text, and split from the publisher training partition. The dataset's attribution and change notice are in the source repository.

The pinned NVIDIA model card's metadata links the June 2024 Open Model License, while its body links the August 2024 Community Model License. Both permit derivative-model distribution, with different conditions. Both exact texts and a NOTICE are included. This research release makes no claim about permission for production deployment. NVIDIA's terms are separate from the source repository's MIT software license; neither agreement is an open-source software license. Review the applicable terms before using or redistributing the adapter.

Integrity

The adapter was selected before the official test. adapter_model.safetensors is 92,325,848 bytes, SHA-256 047303c0755892431269f3fc9c02a6327d08f3ae51b7c495936a3ee2a646bfdb. The original adapter_config.json is SHA-256 7263d1efaeffbaf4b65370ac9c593168a818e3420dfee4f947ed3bf529a2906d. SHA256SUMS.txt covers the packaged files.