Post-Training NVIDIA Nemotron 3.5 Super VL for Finance: More Accurate, Cost-Efficient Financial Credit Analysis

In finance domains like credit investing, the work combines great complexity with massive scale, so finding the right balance between model intelligence, token efficiency, and cost is what unlocks broad coverage of the credit universe. For these use cases, the best model isn’t necessarily the one with the highest benchmark score, but rather the one that balances efficiency and sits at the right point on the Pareto frontier of cost and accuracy.

To test that idea, AfterQuery and Hebbia post-trained the NVIDIA Nemotron 3.5 Super VL open model on expert-written public credit tasks. Perfect solves on a 100-task held-out set rose from 47% to 74%, putting it within three points of GPT 6 Astra out of the box (77%), at roughly one-seventh of the cost (estimated $3.63 for Nemotron to run the 100 tasks against $24.72 for Astra).

The above curve plots each model’s perfect-solve rate (the share of tasks where every rubric criterion was met) against what it costs to run a held-out set of 100 agentic credit-workflow tasks. Post-training on expert-written tasks lifted Nemotron 3.5 Super VL from 47% to 74%, which moves it above the frontier formed by the out-of-the-box models: it now solves more tasks than GPT 6 Sol (63%) and comes within three points of GPT 6 Astra (77%) at roughly one-seventh of the cost. Fable 5.1 remains the most accurate model at 85%, at about 14x the cost. Each model was run once per task on tasks held out from training, and a judge model graded every answer against an expert-written rubric. Scores exclude tasks with no recorded reward (LLM judge errors), 3 to 8 per model. Cost is the solver spend for the full 100-task run, excluding grading, and the two Nemotron cost figures are estimates based on token counts and OpenRouter list pricing.

Those results point to a practical approach: matching models to tasks. Credit work runs at high volume, and cost per document matters as much as accuracy. A post-trained open model, like Nemotron, has near-frontier accuracy at a fraction of the cost, making it a strong fit for high-volume, well-scoped work that makes up most credit workflows. Nemotron also has a strong foundation in math, which makes it well suited to post-training for finance. Teams can still route edge cases to a frontier model when they want an added layer of review.

The efficiency tradeoff pays off most in credit, where a single loan can be governed by a series of documents and amendments, and a portfolio can mean thousands of documents. An open model lets a team post-train once on its own domain, keep the weights, and cover more documents, faster, for less.

“Hebbia works with some of the world’s largest credit investors and providers of credit insights. We’ve seen firsthand how much time and expertise goes into analyzing credit, and the opportunity to make that process faster and more cost-effective. The real value comes from applying that analysis across the entire credit universe, not just individual documents. These results show we can bring credit expertise to tens of millions of documents at a cost that makes it practical.” – Aditya Ramanathan, Head of Applied Research at Hebbia

Focusing on credit workflows was a deliberate choice. The backdrop is a large market: in the US, leveraged loans, the broadly syndicated market these tasks draw on, reached a record $1.6 trillion outstanding at the end of 2025, with private credit managing about $1.4 trillion as of 2025.

The work is also hard for models to parse reliably. Hebbia provided guidance on complex credit workflows that practitioners handle on the job and that models often struggle with. These workflows can run on long, heavily amended credit agreements. Reading them correctly is everyday work for practitioners and a common failure point for models, which have to reason over a large context that mixes quantitative and qualitative information.

To capture that complexity, the credit tasks used to train our model cover economic terms (pricing, margins, floors, leverage), non-economic terms (covenants, consents, governance, default), and analysis that combines the two.

Getting to these results started with a simple question: what makes a credit task hard for a model?

That is the question AfterQuery was built to answer. Our founding team comes from investment banking, private equity, hedge funds, and quantitative trading, and we have built finance training data and benchmarks for frontier AI labs. That background made us a natural thought partner for Hebbia in designing high-quality credit tasks.

Hebbia is an AI platform built for the world’s leading institutional investors. Together, Hebbia’s customers manage over $30T in AUM and use the platform to originate, diligence, and monitor deals. Hebbia also has a deep bench of former credit investors from public credit and distressed debt, which is why it could identify nuanced workflows, such as determining which amendment governs and applying defined terms and exceptions correctly. AfterQuery turned that domain knowledge into tasks written by our in-house credit subject-matter experts.

To keep the tasks focused on one core finance capability, Hebbia encouraged the AfterQuery team to build them around syndicated loan agreements. From there, we moved past simple retrieval and extraction. The tasks center on cases where:

• the operative answer depends on a later amendment
• a definition sits several sections away
• an exception’s scope is easy to miss
• the numerator differs between the maintenance covenant and the pricing grid

Our goal was to see whether targeted post-training could move a baseline Nemotron model closer to frontier performance at a fraction of the cost, or at least deliver a measurable improvement over the baseline.

Hebbia guided towards practical credit tasks, written and reviewed by former credit investors, used for post-training. Hebbia supplied the credit-workflow taxonomy, ontology, representative questions, and domain review. AfterQuery ran collection, task quality checks, answer keys and rubrics, model evaluation, and the post-training study with Nemotron 3.5 Super VL. Post-training was run with NVIDIA NeMo RL open library on 32 H200 GPUs.

Each task comes with a document set: a base credit agreement and several related documents, often amendments. A document set can be a single base agreement or as many as 10 documents spanning multiple amendments with revised language.

The work is moved in four stages:

1. Seed tasks. Hebbia’s credit experts with public credit and complex distressed debt investing experience, wrote initial tasks grounded in real workflows.
2. Expert authoring. AfterQuery’s expert network wrote new tasks modeled on those seeds. A representative profile of an example credit expert would be a professional with two years experience in investment banking and a current, tenured role at a senior associate or principal level as a private or public credit investor.
3. Calibration. Hebbia reviewed a sample to validate the difficulty and practicality of the task.
4. Full review. AfterQuery applied that guidance to review every task in the set.

Each task pairs a real credit agreement, and usually its amendments, with a question a client or an internal stakeholder would ask. The model must reason from the supplied documents (HTM files) and produce a final answer that cites the supporting text. The answer depends on using the right document version, date, defined term, proviso or carve-out.

Each finished task also has a reference answer with citation and rubric criteria. Each rubric follows a few rules. It breaks the answer into atomic requirements, avoids double counting, doesn’t reward facts the question didn’t ask for, and accepts alternative wording when the documents support it. The task review surface examines the question, answer, individual criteria, source support, and domain feedback.

AfterQuery defined proper difficulty by running baseline Nemotron on each task five times and kept only the tasks it solved in three or fewer of the five runs. Examples appear in Learnings from Post-Training below.

Tasks are built around common things a general-purpose AI usually answers incorrectly. Each task targets a point where the agreement differs from market-standard terms, so an experienced credit reader answering from memory would get it wrong. It asks a question a client or an internal stakeholder would ask and supplies every governing document. To answer, the model must read the clause together with its definitions, amendments, and related provisions. As in real-world credit work, the question does not point to the deciding exception or ask the model to simply list a definition.

To confirm a genuine model failure, we first established the correct answer from the version of the agreement in force. We then identified the wrong answer a model is likely to reach by relying on a superseded amendment, a clause read without its proviso, or a nearby covenant that looks applicable. A complete answer has to cite the controlling text and explain why the other provision does not govern.

These details change real credit decisions. A covenant test, pricing margin, guarantee, collateral requirement, or overadvance charge can change when an amendment replaces a definition, a status change switches the governing ratio, or a proviso limits an amount or timing. An analyst who relies on market practice or stops at the first search result can give a confident but incorrect recommendation.

We reviewed Nemotron’s failed attempts task by task. In most failures, Nemotron found the governing clause but left a required qualifier, proviso, second part of the question, or citation out of its final answer. In other failures, it stopped at the first plausible search result, or read the right rule and applied it too broadly or too narrowly. The best examples show where a plausible answer differs from the governing documents and how that difference changes the credit recommendation. Each graded detail should matter to the question asked, so the example shows a real workflow failure and not a formatting miss.

Why customize Nemotron? Before training, Nemotron usually found the right passage but left part of the answer out. The governing clause (i.e., the contract language that decides the answer), was on screen in 99% of runs, but the final answer still left out a qualifier (a condition attached to the rule), an exception, or half of a two-part question. The baseline model showed the same pattern on the held-out task set of 100. Its median answer was about 1,000 characters with one section citation, and three answers were just “7”, “45” and “A”. This matters in credit work because the general rule often isn’t the full answer, which often depends on a proviso, a carve-out, or an amendment.

How post-training customized it for credit analysis. After post-training on expert-written tasks, the model gives complete, cited answers. Perfect solves on the held-out set rose from 47% to 74%, surpassing GPT 6 Sol (63%) and coming within three points of GPT 6 Astra (77%) out of the box, at roughly one-seventh the cost (an estimated $3.63 for Nemotron to run the 100 tasks against $24.72 for Astra). Additionally, the number of grading criteria failed for missing content fell from 58 to 13. The median answer grew to about 4,600 characters, and section citations per answer rose from 1 to 10. The model now quotes the governing clause instead of paraphrasing it, so its answers include which party is covered, the dollar amount, and the exception. It searches slightly more and reads about the same amount of text. The main change is that it now includes what it finds in its answer.

What improvement looks like in practice. Below, the base model’s answer is shown beside the post-trained model’s on the same held-out task, for three representative tasks. Each model ran each task once. The answer text is shown as the model wrote it, with […] marking where a long answer is shortened.

Nemotron 3.5 Super VL — Task asks whether a proposed restricted payment is permitted under an amended credit agreement.

Task excerpt. “In December 2026, Holdings proposes to make a $1,800,000 Restricted Payment that fits within no clause of Section 7.6 other than clause (g). No Default or Event of Default has occurred or would result, and no Restricted Payment has previously been made under clause (g). Giving pro forma effect to the payment, the Consolidated Net Leverage Ratio would be 3.25 to 1.00. Is the payment permitted under Section 7.6(g)?”

Nemotron 3.5 Super VL — Task asks for the interest owed on two advances, one of them repaid the day it was made.

Task excerpt. “Bank makes two Advances to Borrower, each on a Business Day. Advance 1 ($2,000,000) is made on Wednesday, April 27, 2022 (a Funding Date) and repaid in full on Wednesday, May 4, 2022. Advance 2 ($3,600,000) is made and repaid in full on Friday, May 6, 2022. Each bears interest at 5.50% per annum and no Event of Default exists. (1) For how many days does interest accrue on Advance 1? (2) What interest, in dollars, accrues on Advance 2?”

Nemotron 3.5 Super VL — Task asks when interest is payable and which interest periods a borrower may elect after the move from Eurodollar to SOFR.

Task excerpt. “A desk booked quarterly Interest Payment Dates and only one- or three-month SOFR Interest Periods, and also assumed a two-month tenor remains available from the old Eurodollar menu. Under the governing definitions, on which day is each Interest Payment Date, and which Interest Period lengths may the Borrower elect for a SOFR Borrowing?”

Why that matters in a credit workflow. Each improvement affects a decision a credit team makes. The three examples above decide whether a payment is permitted, how much interest is owed, and which interest periods a borrower can elect. In another held-out task, the trained model found that a prepayment notice can be revoked if the refinancing falls through.

Get in touch here to work with AfterQuery on post-training Nemotron open models for your domain-specific use cases, or reach out to us directly at research@afterquery.com.

AfterQuery is an applied research lab curating data solutions to accelerate foundation model development.