went through the HF DS loop in may 2026, sharing specifics because i couldn't find anything useful before mine.
recruiter screen: 30 min, pretty standard. they want to know if you can talk about ML projects without sounding like you copy-pasted from a paper. they mentioned the team i'd be supporting was focused on model evaluation metrics and dataset quality, so calibrate accordingly.
round 1: SQL + stats screen (60 min) this was the one i was most worried about. two SQL problems: one was a window function thing (rank users by engagement within a rolling 7-day window), another involved a self-join to find model runs where a fine-tuned version outperformed the base on at least 3 out of 5 benchmarks. both were medium difficulty, closer to analytics SQL than leetcode SQL. they care that you can write it cleanly and explain the business intuition.
stats portion: a/b test question about detecting if a new tokenizer reduced perplexity significantly. they asked me to walk through the test design, what sample size i'd need, and how i'd handle variance. not super deep, but they wanted to see you could reason about experiment validity.
round 2: case study (90 min take-home + 30 min debrief) they gave me a synthetic dataset of model benchmark scores across 50 models on 6 tasks. asked me to: identify which benchmarks were most predictive of user satisfaction, flag any data quality issues, and propose a scoring methodology. i used pandas + scipy for analysis, built a few visualizations. the debrief was mostly "why this metric and not that one" type questions. felt conversational.
round 3: cross-functional behavioral (45 min) one panel with an EM and a senior DS. standard behavioral but tuned toward how you collaborate with researchers. they really want to know if you can translate ambiguous research outputs into something actionable. i got "tell me about a time you pushed back on a metric choice" and "how do you handle it when a model improves on benchmarks but worse in production."
overall: two weeks from first screen to offer. they moved faster than i expected for a company that size. offer was competitive for a remote-first role. ask me anything.