Nemotron 3.5 Super VL's Open Post-Training Datasets: What You Can Actually Fine-Tune On

NVIDIA Nemotron 3.5 Super VL put its weights on Hugging Face on 12 October 2026 and was announced on 15 October. The half of that release most coverage skipped past is the one you can actually build on: between 12 and 15 October, the nvidia organisation published eleven new post-training datasets — the supervised-fine-tuning and reinforcement-learning corpora behind the model card's own six-stage training description. Every one of the eleven names OpenMDW-1.1 as its governing terms, but that is where the uniformity stops. Two of them present as license: other on the Hub; two others carry a second licence because of what is inside them; and the two cards that never say "ready for commercial or non-commercial uses" are not the ones you would guess. Below is what is in each set, what the cards do and do not publish about it, and how the whole drop lines up against a 99.62% median accuracy-recovery figure that reads very differently depending on which NVIDIA document you open.

What landed, and when

Hugging Face records the creation timestamps, so this is read rather than estimated. All eleven repositories appeared within a 61-hour window:

• Nemotron-RL-litmus-bench-v0.2 — 12 October, 16:37 UTC. Chemistry reasoning, 47,118 train and 2,056 test questions, 64 MB, built from ChEMBL, the OpenADMET ExpansionRx challenge data and a Biogen ADME set. Dual-licensed CC-BY-SA-4.0 and OpenMDW-1.1, because the ChEMBL-derived questions carry the share-alike term.

• Nemotron-SFT-Agentic-v4 — 14 October, 17:54 UTC. Single- and multi-turn tool-use trajectories, 634,664 conversation records across a tool_calling split of 549,707 and a knowledge_agent split of 84,957, 18.88 GB. Generated with Qwen3.5-397B-A17B and GLM-5.2.

• Nemotron-SFT-Instruction-Following-Chat-v4 — 14 October, 17:55 UTC. Chat, multi-turn instruction following and response-length control, 3,850,582 conversation records, 80.74 GB. Generated with Kimi-K2.6, Gemma4-E4B-it and Qwen3.5-397B-A17B.

• Nemotron-SFT-Agentic-Finance-v1 — 14 October, 18:54 UTC. 7,878 multi-turn tool-calling examples across numerical reasoning, financial modelling, market analysis, adjustments and beat-or-miss questions over public SEC filings, 28 MB.

• Nemotron-RL-Agentic-Finance-v1 — 14 October, 18:55 UTC. 10,185 RL tasks, same five task families, 166 MB, with the filing context deliberately withheld so the agent has to retrieve it with sec_filing_search, parse_html_page and retrieve_information.

• Nemotron-SFT-Long-Context-Transduction-v1 — 14 October, 19:47 UTC. 10,000 instruction–answer pairs, 1.22 GB, for position tracking and sequence transformation. No model generated these: the card states explicitly that no language model produces the examples or answers, and the four task families are arithmetic evaluation, hexadecimal-token sorting, variable expansion and CSV transformation.

• Nemotron-SFT-Long-Context-Translation-PG19-v1 — 14 October, 20:14 UTC. 4,335 English-to-multilingual translation examples, 2.22 GB, source texts drawn from public-domain books in the PG-19 corpus and translated with DeepSeek V4 Flash.

• Nemotron-RL-Finance-Tool-Calling-v1 — 14 October, 22:46 UTC. 90,307 finance records — 81,397 core reasoning samples plus 8,910 broader finance-agent samples — in 1.23 GB, driven against the Finance SEC Search environment in NeMo Gym.

• Nemotron-SFT-Finance-Document-Grounded-Reasoning-v1 — 14 October, 22:46 UTC. 63,056 examples in 1.01 GB, grounded in 10-K, 10-Q and 8-K filings from S&P 500 companies covering 2017–2024.

• Nemotron-RL-Agentic-Conversational-Tool-Use-Seed-v1 — 14 October, 23:45 UTC. 100,000 records, 3.86 GB: seed policies, tool definitions and scenarios for the conversation_tool_use_simulation resource server, covering hospitality, e-commerce, banking and insurance, for simulated agentic conversations during RL.

• Nemotron-SFT-BBQ-v1 — 15 October, 05:22 UTC. English fairness data covering age, physical appearance and disability — 2,000 records each, 6,000 in total, 8.0 MB — built against the Bias Benchmark for QA taxonomy, with conversations between fictional participants designed to reward evidence-based answers over assumptions.

Two things about that list matter more than the row counts. The first is that the licence surface is not as uniform as the headline tag suggests, and the second is whose outputs these are.

The licence tag is not the whole licence

All eleven repositories name OpenMDW-1.1 as their governing terms. What differs is how that reaches you, and the differences are worth reading before you build anything commercial on top.

• Tag versus named licence — nine of the eleven carry license: openmdw-1.1 as a plain tag you can read straight off the API. Nemotron-SFT-Agentic-Finance-v1 and Nemotron-RL-Agentic-Finance-v1 present as license: other with license_name: openmdw-1.1 and a link to the agreement. The terms are the same; a tool that filters repositories on a licence tag will see two different values.

• Second licence because of upstream data — Nemotron-SFT-Long-Context-Translation-PG19-v1 carries Apache-2.0 alongside OpenMDW-1.1 because the source corpus is PG-19, and its card adds that the underlying English works are public domain in the United States but that copyright status may differ in other jurisdictions, so users should check local law before redistribution. Nemotron-RL-litmus-bench-v0.2 carries CC-BY-SA-4.0 alongside OpenMDW-1.1, with the share-alike term applied to the ChEMBL-derived questions specifically.

• The commercial sentence — nine cards state "This dataset is ready for commercial or non-commercial uses." The two that do not are Nemotron-SFT-Long-Context-Transduction-v1 and Nemotron-SFT-Long-Context-Translation-PG19-v1 — the long-context pair, not the finance sets.

• Where the content carve-out is documented — finance. The document-grounded card states that the release retains only permitted limited-context excerpts from the bodies of Forms 10-K, 10-Q and 8-K, and that the full SEC corpus, all exhibits and any content bearing a third-party copyright notice are excluded, naming consultant reports, presentations, fairness opinions and analyst research. It also notes that excerpts remain the responsibility of their respective issuers.

So the practical reading is the opposite of the intuitive one. The finance sets are the ones with the loudest licence badge — a clean openmdw-1.1 on the document-grounded and tool-calling sets — and the ones with the most documented restrictions on what is inside them. The long-context sets are the ones whose Hub metadata and prose disagree most with each other. Neither is a reason to walk away; both are reasons to read four cards instead of one license field.

Whose outputs trained NVIDIA's flagship

The other thing that runs across the drop is the generation models, and the cards name them without hedging. The flagship agentic SFT set was produced with Qwen3.5-397B-A17B and GLM-5.2. The instruction-following set adds Kimi-K2.6 and Gemma4-E4B-it. The document-grounded finance set names Qwen3.5-397B-A17B for question generation, DeepSeek-V4-Pro for answers and gpt-oss-120b for filtering and selection. The RL finance tool-calling set lists Qwen3.5-397B-A17B and Qwen3-235B-A22B-Instruct-2507 upstream, DeepSeek-V4-Pro for answers, gpt-oss-120b for filtering, and NVIDIA Nemotron itself for RL difficulty profiling across eight rollouts. The PG-19 translations were produced with DeepSeek V4 Flash.

That is a publishing decision, not an accident, and it is the most useful thing in the drop for anyone trying to reason about quality. It tells you the behavioural prior you are fine-tuning into is substantially not NVIDIA's — it is a blend of two Chinese frontier open-weight families, an OpenAI open-weight model and a Google small model. If you have an opinion about how any of those behave on financial text or on tool selection, that opinion applies to your starting point here, and you now have the provenance to check it rather than guess.

Why the data is the part you can act on

The reason a dataset drop is the interesting half of this launch is visible in the vendor's own numbers. NVIDIA's newsroom post names organisations that post-trained the model and quotes their results: Scale AI for Skydance's vendor credit work, at "40x lower inference cost" and "5x lower latency"; Ramp on converting receipt images into structured JSON, expecting "up to 50% higher throughput and 10x lower latency"; AI21 and Alice on threat intelligence, at "41% higher accuracy." Read the developer blog alongside the newsroom and the same customer results collapse into "40x lower inference cost, 10x lower latency, and 40% higher accuracy" — a 40 where the newsroom says 41, and a 10x where it says 5x, for what look like the same two deployments. These are customer expectations for their own workloads, not benchmarks, and neither version is reproducible by you.

What is reproducible is the receipt. The newsroom says the model was post-trained with roughly 3.8 million samples and reinforcement learning across dozens of environments, and that the open datasets span agentic tool use, finance, instruction following, long context, fairness and vision-language understanding. The eleven repositories above are that corpus, and the NeMo Gym environments the cards point at are the harness. A team that wants a document-intelligence agent now has the model, the serving recipes, the fine-tuning libraries and the training data in one place. The gap between "this model can be specialised" and "here is a specialisation you can download today" is what these repositories close.

One caveat on that list of categories: NVIDIA's newsroom includes vision-language understanding among the released datasets, and the eleven new nvidia dataset repositories created in this window are text and tool-use corpora. The document-grounded finance set is the closest thing to a vision set in the drop, and it is text-only in format. If you came for the image and video side, the data is not here — the most recent NVIDIA vision dataset on the Hub predates this window by five days.

What the cards do not publish, and what that costs you

Here is the number nobody puts in a launch post, and on this drop it is a missing number. None of the eleven dataset cards publishes a token count. They publish record counts and storage sizes, and that is it. So the sizing question — how many GPU-hours is one epoch over the instruction-following set — cannot be answered from the release, only estimated by dividing bytes by an assumed bytes-per-token ratio, which is exactly the kind of figure that travels into coverage as if it were measured.

What you can work from is the published storage: 80.74 GB for Nemotron-SFT-Instruction-Following-Chat-v4, 18.88 GB for Nemotron-SFT-Agentic-v4, 3.86 GB for the conversational tool-use seed set, and then a long tail — 2.22 GB, 1.23 GB, 1.22 GB, 1.01 GB, 166 MB, 28 MB, 8.0 MB, 64 MB. The two largest sets are one to two orders of magnitude bigger than everything else in the drop, which tells you where the post-training weight actually sits without needing a token count at all.

• Nemotron-SFT-Instruction-Following-Chat-v4 — 80.74 GB and 3,850,582 records. This is the set that makes a full-parameter run a project rather than an afternoon.

• Nemotron-SFT-Agentic-v4 — 18.88 GB and 634,664 records. Roughly a quarter of the size, which puts it in reach for a single-node run.

• Nemotron-SFT-BBQ-v1 — 8.0 MB and 6,000 records. Small enough to iterate on in minutes, which makes it the cheapest way to prove your fine-tuning loop is wired up before you commit a GPU-week to the job you actually care about.

• Nemotron-SFT-Long-Context-Transduction-v1 — 1.22 GB, but only 10,000 examples, with a published per-example range of 2,048 to 131,072 target tokens. Here the cost is driven by sequence length rather than volume, and the card says the budgets are generation targets estimated with cl100k_base, not measured post-training counts.

For scale on the model side, the model card publishes the stage budgets: a 100-billion-token MTP warm-up against a frozen backbone, one trillion tokens of distillation from the next-larger Nemotron, 100 billion tokens of Indic continued pre-training, a 35-billion-token extension to a 1M context, 67 billion tokens of software-engineering mid-training at 256K, and a 9-billion-token long-context recovery stage. Those are the only token budgets NVIDIA states anywhere in the release, and they apply to training the base model, not to fine-tuning on these datasets.

That is also why the LoRA path matters here rather than being a footnote. NVIDIA points developers at NVIDIA NeMo AutoModel for supervised fine-tuning and low-rank adaptation, and adapters are the difference between moving a few of the model's behaviours and re-deriving 249 GB of weights. Both paths are real, and the choice between them is a decision about how much of the base model's behaviour you intend to move. A finance team chasing a narrow schema or a policy-following behaviour is in LoRA territory. A team that wants the model to hold a new operating procedure across every one of its interactions is not.

Two documents, one benchmark, opposite answers

The technical report accompanying this release is dated 14 October and carries the number that has travelled furthest without its context. NVIDIA reports that the released NVFP4 checkpoint achieves 99.62% median accuracy recovery relative to BF16 across 36 benchmarks, after the aggressive post-training-quantization recipe followed by 200 iterations of SFT-based quantization-aware distillation — with the C-RADIO vision encoder, the VL connector, the Mamba convolutions and the embeddings left in BF16 and the router computation left in FP32, per the report's per-module table.

"Median accuracy recovery" is a median of ratios, not a minimum. To get 99.62% the report computes a score ratio per benchmark and takes the middle one across 36, so the headline tells you the typical benchmark lands about four-tenths of a percent below full precision, and nothing at all about the tail. The report's own table shows the tail: Tau3 Banking at 94.20%, SciCode's problem-solving split at 93.63%, MMLU-ProX at 95.90% — and in the other direction GDPVal at 111.53% and Terminal-Bench v2.1 at 103.47%, because quantisation can land above the BF16 reference on a single run.

Then compare the model card, and the story gets more interesting than a tail. NVIDIA's own NVFP4 model card publishes a BF16-versus-NVFP4 table for the same checkpoint, and on it Terminal-Bench 2.1 reads 66.70 against 60.80 — the quantised build nine per cent down. In the technical report's full-suite table, the same benchmark reads 60.67 against 62.78, with the quantised build ahead. Same model, same benchmark, same eval harness family, two NVIDIA documents, opposite directions. The card also shows WebVoyager at 77.72 falling to 66.31 and DeepSWE at 24.65 falling to 18.2, neither of which is the version of the story you get from a 99.62% headline.

That is not a scandal — different tables can carry different harness configurations and the report's table reports a recovery percentage alongside each row — but it does mean the honest answer to "does NVFP4 cost me accuracy" is: on most published rows, barely; on the rows where it does, by more than the median implies; and the vendor's two documents disagree about which rows those are. If your workload is agentic terminal work, re-measure it yourself rather than picking a table.

The deployment arithmetic, by contrast, is unambiguous and is the strongest argument for the quantised build. The BF16 repository totals 248.9 GB of checkpoint files and its own quick-start serves it on 4× H100. The NVFP4 repository totals 75.06 GB, which the report puts at 4.82 bits per parameter including scales, and that fits one H200, one GB200 or a single DGX Spark. On one GB200 the report quotes up to 4,856 output tokens/s/GPU on a 20-turn agentic proxy workload while holding at least 50 tokens/s/user, with NVFP4 at 2.43× the throughput of BF16 on the same workload, and the model at 2.34× and 1.90× the throughput of Qwen3.5-122B-A10B and Qwen3.8 Flash-Next. All of those are NVIDIA-measured, on NVIDIA's own serving stack, at whatever concurrency the recipe chose.

One further quarter of the release is easy to miss and matters if your fine-tune is going to be argued about later. The report describes a contamination pass over the post-training data that removed 20,671 samples — about 0.09%, split into 14,793 SFT and 5,878 RL samples — matching 1,808 benchmark questions, or 6.0%, in 22 of the 30 benchmarks checked, with roughly 3% of the judged candidate pairs flagged as severe contamination. Most open-weights releases publish no cleaning methodology at all. Having one you can cite is worth more than it sounds when a customer asks why your fine-tune's eval numbers look the way they do.

How this changes what you can do this week

Three concrete things are now possible that were not before 12 October.

The first is a finance filing agent you can build without a data-licensing conversation — with one caveat attached. The finance repositories cover the full loop: 63,056 document-grounded SFT examples over seven years of S&P 500 filings, 90,307 RL tool-calling records against a live SEC search environment, 7,878 multi-turn tool-calling examples for the behavioural layer, and 10,185 RL tasks with the filing context withheld so retrieval is part of the task. The caveat is the carve-out documented above: excerpts only, exhibits excluded, third-party-copyright material stripped, and issues remaining the issuers' responsibility. That is a well-documented position rather than a clean one, and it is the set where you should read the card before the code.

The second is agentic tool use, and this is the best-supported vertical in the drop. Between Nemotron-SFT-Agentic-v4's 634,664 records, the 100,000-record conversational tool-use seed set with its hospitality, e-commerce, banking and insurance scenarios, and the NeMo Gym environments the cards point at, the whole pipeline is described end to end: generate trajectories, run simulated conversations, train with verifiable rewards, evaluate in the harness the vendor itself used. Whatever you think of the model, that is a complete public recipe for turning a general vision-language model into a tool-calling agent, and it is the part that used to be tribal knowledge.

The third is a fairness baseline you can actually iterate on. Nemotron-SFT-BBQ-v1 is 6,000 records and 8.0 MB, it is balanced 2,000/2,000/2,000 across age, physical appearance and disability, and its conversations are written so that the evidence-based answer and the assumption-based answer are both plausible. At that size you can run it, score it and re-run it in an afternoon, which turns "how did you test for this" from a procurement problem into an experiment.

Trying it without betting a production path on it

The awkward part of this release is deployment, not data. The BF16 checkpoint wants four H100s and the quantised one wants a single H200 or GB200, the open weights are on Hugging Face, and the vendor's own route to a hosted endpoint is NVIDIA NIM, with the newsroom also naming third-party platforms. We do not serve this model, so if you want a managed endpoint the vendor's own API and several third-party platforms are where to look; what a router is good for here is the rest of the decision.

One key across 200-plus models at provider list price with 0% markup means comparing the NVFP4 build against whatever text model you run today costs a single line change rather than a second vendor contract, and automatic failover across providers means a workload you are only testing can sit next to production traffic without either one depending on the other. That is the shape of a decision you can reverse — which is what you want from a model whose documentation is two days old and whose own benchmark tables disagree with each other.

The practical sequence is: take Nemotron-SFT-BBQ-v1 or a slice of the agentic set, run a LoRA on a rented H200 to prove your loop, score it against your own eval, and only then decide whether the 249 GB full-precision build and four GPUs are worth it. The datasets are free to download and the small ones are free to run. Nothing about that plan requires you to believe a 99.62% median.

What is still missing

Three gaps are worth watching because they change the calculus rather than the marketing.

The first is independent evaluation of the fine-tuned results. Every post-training number in this release — the 62.7-to-71.9 Terminal-Bench move from parallel agent execution and self-verification, the 24.6-to-34.5 DeepSWE move in the same experiment, the 41% accuracy improvement at AI21, the 40× cost reduction at Scale AI — comes from NVIDIA's harness or from a customer's own measurement. There is no third party that has reproduced a Nemotron 3.5 Super VL fine-tune on these datasets and published the result. Until there is, treat the effect sizes as an upper bound.

The second is the video path. The model accepts MP4 up to two minutes at 2 FPS and 256 frames, or 1 FPS and 128 frames for 1080p, with a pruning rate of 0.5 — and there is no public video post-training dataset in this drop. Every repository above is text, and the translation and transduction sets are long-context text rather than anything visual. If your use case is the video understanding the model card advertises, NVIDIA has shipped the capability without the data to specialise it.

The third is that nobody has published what these datasets do to a model that is not Nemotron. All eleven were built to produce one model's post-training mixture — the cards are explicit that some sets consolidate earlier Nemotron-specific versions, and the conversational seed set is described as the seed data for a pivot set rather than fresh generation. Whether the agentic corpus transfers to a different 120B backbone is an open question, and it is the first one a team with an existing fine-tuning stack will want answered.

All three gaps are the kind that close with someone else's work rather than NVIDIA's, which is the usual arc for an open release. The weights are downloadable today, the data is downloadable today, and the first credible independent fine-tune is the thing that will tell you whether any of it is worth the GPU time.