PhD Internship, NYC, In-Person

The Situation

At Plastic Labs we are building Honcho, a continual learning system for modeling personal identity. Our reasoning models learn to understand people, agents, and groups as their preferences, needs, and circumstances change. Building our training and evaluation datasets requires understanding how conversations differ across settings, how they vary within a use case, and which differences matter when selecting examples.

You'll work with our ML team to conduct these analyses and use the findings to improve our data preparation, evaluation, and human labeling workflows.

The Job

You'll contribute datasets and methods to our ongoing training and evaluation work, taking your findings from analysis through implementation.

Understand the data. Analyze the tasks and conversation patterns in our datasets. Establish what is common, what is rare, and where missing context limits an example's usefulness.

Design the stratification. Define categories that capture relevant differences between examples, drawing on characteristics such as use case, conversation length, language, and number of participants.

Define what to keep. Develop cleaning and deduplication methods that distinguish unusable records from difficult but valuable examples. Inspect what each rule removes, looking for disproportionate exclusions of particular categories or unintended shifts in the distribution.

Design the samples. Build training datasets, evaluation sets, and human labeling batches with a sampling approach suited to each. Account for relationships between examples when creating splits to prevent overlap from biasing evaluation results.

Test your choices. Compare automated labels with human judgments and investigate disagreements. Measure how cleaning or sampling choices affect the conclusions drawn from a dataset, using those results to recommend what to train on, measure, or investigate next.

Make it repeatable. Deliver versioned datasets with the code used to produce them. Document the reasoning behind your choices so the team can reproduce the results and apply your methods to new data.

You

• You're pursuing a PhD in machine learning, statistics, NLP, computer science, or a related discipline.
• You've worked closely with messy, real-world data. You can explain how choices about collection, filtering, and sampling affect a research result.
• You have strong statistical judgment. You can explain what an analysis supports, where it may be biased, and how to validate its conclusions.
• You can work through an open-ended problem. You propose an approach, implement it, and revise it when the evidence disagrees.
• You write clear analyses and usable code. Someone else can reproduce your results and understand your assumptions.

Technical Requirements

• Strong Python
• Comfortable with SQL
• Experience handling large datasets

Nice to Haves

• You've used text analysis, clustering, or human annotation to organize conversational data or check labels produced by language models.
• You're interested in memory, personal identity, or interactions between humans and agents.

Apply

Send us two things.

1. A research project or analysis where your decisions about data mattered. Tell us what you found, how you checked it, and what you'd change with more time.
2. Your CV and any relevant papers, code, or other work, along with your availability.

APPLY HERE

(Back to Working at Plastic)