Member of Technical Staff, Pretraining Data

Mirendil Inc.; San Francisco (CA); In loco; USD 300.000 - 400.000; Tempo pieno

Descrizione del lavoro

Mirendil is building frontier AI research capabilities and scaling its own models. You will set the data strategy for pretraining, design ablations, and direct agents to run experiments with researchers who own architecture and training runs.

You will extend data pipelines, improve data extraction and deduplication, and develop data scaling plans for each run. Strong Python, Spark, PyTorch/JAX experience and clear written reasoning are highly valued.

Competenze

• Proven experience shipping data changes in large pretraining runs.
• Experience building data pipelines at trillions of tokens with Spark, Ray, Beam or similar.
• Ability to run training experiments using PyTorch or JAX.
• Hands-on experience with deduplication, quality classifiers, mixture optimization, data scaling laws, or long-context and code data.
• Excellent written communication to explain experimental results.

Mansioni

• Set the direction for data used in model training and run ablations and analyses.
• Direct agents to run experiments and pipelines in parallel.
• Design data mixtures and scaling strategies for each run.
• Extend data pipelines into scientific domains and improve data extraction and deduplication.
• Keep corpus provenance and reproducibility, decontamination for evals.
• Collaborate with researchers who own architecture and training runs.

Conoscenze

Python

Data pipelines

Pytorch

JAX

Data quality

Experiment design

Strumenti

Spark

Ray

Beam

Git

Descrizione del lavoro

Mirendil

Mirendil is a tech-first company focused on solving core bottlenecks that unlock step-change acceleration across science and technology. Our first goal is to democratize frontier AI R&D across scientific disciplines. We are building a frontier AI research company and training our own models end-to-end.

The Role

We pretrain our own models and we are scaling to larger models now. Agents handle much of the engineering, but deciding what goes into training still depends on strong human judgment.

You will set the direction for what our models train on and use agents to run the ablations, pipelines, and analyses that test it. You will work closely with the researchers who own architecture and training runs.

What You’ll Do

• Work on building data pipelines for pretraining, including repository-scale code and technical documents.

• Get more out of public data through better extraction, filtering, and deduplication of web pages, open-source code, and public datasets.

• Decide the data mixture for each run, and design the ablations and data scaling experiments that justify it.

• Direct agents to run those experiments and pipelines in parallel, and decide which results to trust.

• Extend a code-first mixture into scientific domains, starting from where our models fall short.

• Keep the corpus clean as it grows: provenance, decontamination against our evals, and reproducible versions.

About You

• You have shipped a data change into a large pretraining run and can show the ablation that justified it.

• You have built data pipelines at the scale of trillions of tokens with Spark, Ray, Beam, or similar, and you write strong Python.

• You know PyTorch or JAX well enough to run your own training experiments.

• You have hands-on experience with at least two of: deduplication, quality classifiers, mixture optimization, data scaling laws, or long-context and code data.

• You can tell when an experimental result is real, and you explain your reasoning clearly in writing.

If you’re excited about shaping what our models learn from, and proving it with experiments, we’d love to hear from you.

We offer a base salary of $300,000–$400,000 USD and a meaningful equity grant, depending on experience and background, along with competitive benefits.