Marah Abdin. Synthetic data researcher. Synthetic data team lead at Poolside; previously Phi at Microsoft.
marahabdin.com
Bio, Work & Ideas
Marah Abdin
Conference affiliation: Team Lead - Synthetic Data · Poolside · 2026
On this page
Marah Abdin leads synthetic-data research at Poolside, developing the training material behind its Laguna language models for coding and agentic software development. Her specialty is turning raw text, code, and mathematics into carefully structured examples that strengthen a model’s reasoning without simply multiplying repetitive training data.
From Microsoft’s Phi models to Poolside
Abdin worked in research and engineering across Microsoft Research’s Redmond and New York laboratories before joining teams at Microsoft Research and Microsoft AI focused on large language models. She contributed to the Phi family from Phi-1.5 through Phi-4-reasoning and coauthored the technical reports for Phi-3, Phi-4, and Phi-4-reasoning. That work explored how curated datasets, synthetic examples, and reinforcement learning can improve compact models’ performance on mathematical and scientific reasoning.
In 2025, she joined Poolside and established its synthetic-data team, spanning pre-training, post-training, reinforcement learning, and agent-based generation. She coauthored the Laguna M.1 and XS.2 technical report and highlighted its synthetic-data methods, including the accumulation of more than 4.4 trillion synthetic tokens.
How she designs training data
Synthetic data complements organic data. Generated examples make implicit reasoning, planning, and structure within naturally occurring material easier for models to learn. Synthetic data comprised 13 percent of Laguna XS.2’s pre-training mixture.
Diversity matters more than repetition. Reusing scarce, high-quality examples can saturate models prematurely. Abdin applies scalable rephrasing and specialized pipelines that convert raw code into code-plus-explanation examples or generate STEM-focused material.
Composable synthetic-data pipelines. Her systems combine seed material, metadata, generators, filters, and validators. Smaller models can handle inexpensive, seed-driven transformations; harder tasks demand staged generation and stronger correctness checks.
Multi-agent orchestration through Hive. The configurable system coordinates agents through an orchestrator and supervisor, enabling mathematical problems to become code, complex tasks to be decomposed, and generated material to improve through iterative interaction. These methods formed Abdin’s portion of her joint AI Engineer conference talk.
