Miju Labs

← All buyers · NVIDIA (Nemotron / Cosmos)

Z

Zhilin Wang

Senior Applied Scientist, NVIDIA (NeMo NLP / LLM alignment) · NVIDIA (Nemotron / Cosmos)

medium confidenceTier A · Talk to firstSecond-wave labstext

The person who actually designs NVIDIA's human preference datasets. He is lead author of HelpSteer, HelpSteer2 and HelpSteer3-Preference and co-author of SteerLM, and he writes the public documentation for them — the annotation scales, the qualification bars per domain, the vendor split between Scale AI and Translated. If Julian wants to know what a frontier buyer's spec looks like, or wants to be measured against one, this is the operational counterpart.

Background

Studied natural language processing at the University of Washington and joined NVIDIA's NeMo NLP team, where his stated goal is making large language models controllable, safe and accessible. His output is a sequence of open alignment datasets and reward-model releases rather than academic prestige papers: SteerLM (2023), HelpSteer (2023), a reward-model release improving alignment with human preferences (2024), Llama Nemotron reasoning models (2025) and HelpSteer3-Preference (2025).

His LinkedIn headline describes him as a senior applied scientist on NVIDIA LLM alignment. Notably he does not appear in the Nemotron 3 (Dec 2025) contributor lists, which may simply reflect a different project assignment rather than a departure.

What they run now

Human preference data and reward models for alignment. HelpSteer3-Preference is his most recent public programme: over 40,000 preference samples across General, STEM, Code and Multilingual, annotated by more than 6,400 people in 77 countries on a seven-point relative-quality scale, with domain-specific qualification bars — a relevant degree for STEM, software-engineering experience for Code, verified fluency for Multilingual — and explicit instructions permitting web search but prohibiting LLM tools. Scale AI supplied General, STEM and Code; Translated supplied Multilingual. He is the named contact on the dataset.

What they buy

Documented, at the spec level: he defines the task, the rating scale, the rater qualification criteria and the quality controls, and NVIDIA procures the labour through Scale AI and Translated. The modality is strictly text and code so far — there is no public evidence of him commissioning image, video or design judgement. Inference: he would be the person a vendor's pilot is graded by, while the contract itself sits above him.

Career

  1. presentSenior Applied Scientist, NeMo NLP / LLM alignment, NVIDIA — Lead author of the HelpSteer dataset family.
  2. —University of Washington, Natural language processing

On the record

Annotator qualification · 2025-05

Sets domain-specific hiring bars for annotators — degrees for STEM, SWE experience for code, fluency for language. source

No LLMs in the loop · 2025-05

Annotators may use web search to verify facts but are prohibited from using LLM tools. source

Controllable models

States his aim as making large language models controllable, safe and accessible. source

Interviews & talks

Why it matters here

Two reasons, only one of them commercial. First, he is the closest thing in this index to a published reference specification for buying expert human judgement; Julian should read HelpSteer3 as a template and be able to talk in its vocabulary with any buyer. Second, he is a realistic technical contact at NVIDIA — but only if Julian can offer something in text-adjacent subjective quality (writing quality, document or chart comprehension, visual reasoning), because design taste is simply not on his roadmap.

How to reach

Speculation: the opener that fits him is a methodological one — how he handles annotator disagreement on subjective dimensions, and whether the seven-point scale holds up for aesthetic or layout judgements. Offer a small openly-releasable pilot built to HelpSteer conventions so it slots into his publishing habit. He is reachable through NVIDIA's developer-blog and dataset channels and responds in public to dataset questions on Hugging Face. Do not pitch design taste cold; pitch method transfer.

What we could not establish
  • Why he is absent from the Nemotron 3 contributor lists — reassignment, or a move within or out of NVIDIA.
  • Whether an X account (@wangzhilin123 appeared in search) is really his; I could not tie it to NVIDIA.
  • Whether NVIDIA's post-2025 preference work still uses Scale AI, and whether any visual preference programme exists.

Why they are on the map

Lead author HelpSteer3-Preference: 6,400+ annotators from 77 countries via Scale AI and Translated source — evidence dated 2025-10

Runs vendor projects; decides if a pilot converts.

How NVIDIA (Nemotron / Cosmos) buys data

HelpSteer3-Preference (2025) used 6,400+ annotators in 77 countries sourced/managed by Scale AI (general, STEM, code with degree/SWE bars) and Translated (multilingual). Nemotron 3 report (Dec 2025) lists dedicated Data, Post-training, Evaluation/Safety and Project Management groups but no annotation vendor; RL is mostly environment-based.

Nemotron post-training under Oleksii Kuchaiev (Applied Research), leadership incl. Jonathan Cohen, Bryan Catanzaro, Kari Briski; HelpSteer human-data programme run by Zhilin Wang and co-authors; 15-person project-management pool. Visual generation in Ming-Yu Liu's research group.

Known vendors: scale-ai, Translated

Sources

  1. NVIDIA Technical Blog — author page for Zhilin Wang (title, bio, posts)
  2. HelpSteer3-Preference (arXiv 2505.11475) — annotation design, vendors, scale
  3. HelpSteer3 dataset card
  4. LinkedIn headline — 'Sr. Applied Scientist @ Nvidia LLM Alignment'