The role
- COMP
- $200K - $350K
- EQUITY
- Competitive Series B equity
- LOCATION
- San Francisco
- WORKPLACE
- On-site
- EXPERIENCE
- 1.5 - 6 years
- VISA
- None, Visa transfers
- STACK
- Python, PyTorch
- INDUSTRY
- AI, Data
/roles — ROLE_287
Research Scientist - Post-Training
Small, venture-backed startup whose datasets and evaluations help leading AI labs improve their models
The company
Frontier-data company capturing how experts reason and turning real professional work into training data for foundation models.
- STAGE
- growth-stage
- FUNDING
- $30M+ raised
- TEAM
- ~80 people
- FOUNDED
- 2024
JD — the work
About the role
Your job would be to prove, with evidence, that the company's data makes models better. You'd set up SFT and RL post-training runs that isolate how particular datasets move alignment, generalization, and capability, then turn the results into clear, defensible claims for partner labs. The work is experimental and high leverage, touching finance, policy, software, and other domains. The team is small and on-site in San Francisco.
What you'll do
- Design tightly controlled post-training experiments, both SFT and RL, that measure what each dataset does for performance
- Quantify improvements in tool use, reasoning, domain-specific workflows, and long-horizon tasks
- Present results to partner labs in support of the sales effort
- Feed your findings back to internal data teams so they can raise quality
What they're looking for
- Strong working knowledge of how LLMs are trained and evaluated
- Real fascination with how data selection, structure, and quality drive what a model does
- Skill at quick, lightweight experiments and at pulling clear lessons from noisy results
- Comfort across many domains and a preference for building over theorizing
- Undergraduate or master's research experience; a PhD is not expected