One Environment, Whole Lifecycle: Agentic Post-Training for Nemotron in Finance

TL;DR

• The artifact being open-sourced is the pipeline, not the weights: the data, the SDG recipes, the training recipes, the environment, and the orchestration are all released so others can regenerate the data and retrain themselves.
• Finance is hard because it demands everything at once — multi-turn reasoning, tool calls, long mixed structured/unstructured filings, financial math, and a citation back to the supporting passage. A question like "calculate the inventory turnover for US Steel in 2024" exercises all five. Today the best closed model sits around 64% and open models around 60%; on a newer, harder finance benchmark the best result is only 58%.
• The load-bearing design choice is one shared environment: SDG, SFT, and RL all run against the same tools in the same format, so consistency is structural rather than a discipline problem. Result: roughly +11% on Qwen models and +7% on Nemotron after fine-tuning, with 25% fewer tokens at no accuracy cost.

Key Points

The goal: releasing the orchestra, not a checkpoint (~02:23–02:25)

Gharghabi's team at NVIDIA works on post-training reasoning models for finance on top of Nemotron. She set the goal out front: release an open-source model that does financial reasoning — but explicitly not just a checkpoint that people load and run. The whole pipeline goes out: open data, the recipe and pipeline used to generate that data, the model, and the environment. Her framing was that they want to share the entire "orchestra" of the process, so that others can generate their own data using the same recipe.

Why finance is genuinely hard: unpacking one inventory-turnover question (~02:25–02:27)

Answering a finance question correctly requires a full multi-turn reasoning trajectory: calling multiple tools, running web search, retrieving different SEC filings. Those filings are long and complex, mixing structured and unstructured content. The model must not only read them but perform financial math over them, and then cite which part of the document supported the answer.

Her worked example: "calculate the inventory turnover for US Steel in 2024." The model has to know the turnover formula; identify which filing and which form it needs, and from which tool (web, SEC); extract and understand the retrieved documents; compute the formula; and cite the supporting evidence so the answer can be confirmed as grounded.

The gap to close is large. On a well-known recently released finance benchmark, the best closed model reaches roughly 64% and the best open-source model roughly 60%. On a newer, more complicated finance benchmark, the best current performance is only 58%.

Four contributions and the numbers (~02:27–02:29)

She grouped the team's work into four pillars:

1. Two synthetic data generation (SDG) pipelines: document-based SDG and template-based SDG.
2. Publicly released data: more than one million samples, available now.
3. Models trained on that data using their own SFT and RL recipes, with substantial measured improvement.
4. RL fine-tuning with tool calling supported natively, all inside the same pipeline.

She also noted two financial-institution customers already using it. Security is a hard requirement for finance data, and the pipeline runs entirely without internet access, which she described as enterprise-ready.

Results: fine-tuning with their data and pipeline gave roughly +11% on Qwen models and +7% on Nemotron models, and — without losing accuracy — 25% lower token usage. Her gloss: "less token, less cost, better life."

The core design: one shared environment (~02:28–02:29)

Everything is orchestrated under NVFlow: synthetic data generation, model fine-tuning, evaluation, and RL training all run through a single pipeline.

The decision she singled out as important is the shared single environment: SDG, SFT, and RL tuning all use the same tools and the same format, so there is nothing to keep manually in sync. That is exactly what the talk title means by "One Environment, Whole Lifecycle."

She closed with an invitation to collaborate — code and data are all public.

Quotes

"It's not just sharing a checkpoint so everyone can load it and run the model — but we want to share the whole pipeline." (~02:24)

The unit of open-sourcing moves from the model to the process.

"All the synthetic data generation, SFT tuning, RL tuning — all of them use the same tools, same format. So no need to be worried about being consistent." (~02:29)

Consistency enforced by architecture rather than by discipline.