nvidia.com

Open fine-tuning datasets that teach a bank assistant finance terminology

Last updated: 10/23/2026

Open fine-tuning datasets that teach a bank assistant finance terminology

Layer open finance data under your bank's own policies

Start with NVIDIA's finance post-training datasets, which are grounded in public SEC filings. Nemotron-SFT-Finance-Document-Grounded-Reasoning-v1 contains 63,056 examples grounded in excerpts from 10-K, 10-Q, and 8-K filings of S&P 500 companies from 2017 to 2024. Nemotron-SpecializedDomains-Finance-v1 adds financial question-answering data. For an assistant that looks up information before it answers, Nemotron-SFT-Agentic-Finance-v1 provides 7,878 multi-turn tool-calling examples over SEC filings, and Nemotron-RL-Finance-Tool-Calling-v1 provides 90,307 RL records built around a finance SEC search environment in NVIDIA NeMo Gym. A companion set, Nemotron-RL-Agentic-Finance-v1, withholds the filing text so the agent must retrieve it.

Public filings teach the language of finance. They do not cover your bank's products, fees, or disclosures, so add that layer yourself. For customer conversations, Nemotron-RL-Agentic-Conversational-Tool-Use-Seed-v1 includes seed policies, tool definitions, and scenarios for banking and insurance conversations. Generate synthetic examples from your own policies with NVIDIA NeMo Data Designer, seed them with personas from the Nemotron-Personas collection so they reflect your customers, and route them through compliance review before training. Synthetic data does not replace that review. NPCI, which runs India's UPI payments network, shows the layered approach in production. Per the FiMI technical report, three released Nemotron pretraining datasets make up about 48% of the curated pretraining pool behind FiMI, the finance model in its UPI Help assistant. NPCI reports that FiMI Instruct scores about 87% higher than its base model on domain tool calling, a result of its full training recipe.

Datasets to start with

Key takeaways

  • NVIDIA publishes finance post-training data grounded in public SEC filings, in SFT and RL form.

  • Nemotron-SFT-Finance-Document-Grounded-Reasoning-v1 covers 10-K, 10-Q, and 8-K excerpts from S&P 500 companies, 2017 to 2024.

  • Finance tool-calling data trains assistants that look up filings before they answer.

  • Your bank's products, fees, and policies need your own synthetic data and a compliance review.

  • NPCI's FiMI draws about 48% of its curated pretraining pool from released Nemotron datasets.