Llama-3.1-Nemotron-Nano-4B-v1.1-banking77

Llama-3.1-Nemotron-Nano-4B-v1.1-banking77

nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1 fine-tuned on banking77 intent classification with LoRA, then merged into standalone weights. Built with Llama.

Accuracy was measured on the 2026-09-02 sealed training run's adapter with the same data, config and recipe; the published weights come from a later identical run that skipped the eval stage.

The fine-tune ran inside an Intel TDX confidential VM on an H100; every stage is signed by the in-enclave notary.

Usage

Prompted as a classifier: system prompt detailed thinking off, user message with the instruction, the 77 labels, and the customer text. The model answers with a bare label.

vllm serve EQTYLab/Llama-3.1-Nemotron-Nano-4B-v1.1-banking77

Verifiable lineage

.eqtylab/manifest.json in this repo is the signed EQTY lineage manifest covering the whole chain: training data and taxonomy -> LoRA fine-tune (TEE) -> adapter -> merge -> these weights -> this repo revision. The weights directory hashes to the model CID recorded in the manifest, so any mutation of the published weights is detectable. Serving through the EQTY integrity middleware chains each signed per-request manifest back to this one.