Staff Research Engineer, Post-training & Evaluation

Remote GREENHOUSE

Reddit · Remote - United States

Job TypeRemote

LocationRemote - United States

Posted2 days ago

Apply Now Redirect to Company Website

AI Summary

Generating summary...

Role DNA

Analyzing role...

Job Description

Reddit is a community of communities. It’s built on shared interests, passion, and trust, and is home to the most open and authentic conversations on the internet. Every day, Reddit users submit, vote, and comment on the topics they care most about. With 100,000+ active communities and approximately 126 million daily active unique visitors, Reddit is one of the internet’s largest sources of information. For more information, visit www.redditinc.com.

Reddit is continuing to grow our teams with the best talent. This role is completely remote friendly within the United States. If you happen to live close to one of our physical office locations (San Francisco, Los Angeles, New York City & Chicago) our doors are open for you to come into the office as often as you'd like.

The AI Engineering team at Reddit is building our own Reddit-native foundational Large Language Models (LLMs). This team sits at the intersection of applied research and massive-scale infrastructure, training models that truly understand the unique culture, language, and structure of Reddit communities. You'll join a team of distinguished engineers and researchers building the "engine room" of Reddit's AI future — the foundational models that power Safety & Moderation, Search, Ads, and the next generation of consumer products.

As a Staff Research Engineer for Post-Training & Evaluation Science, you will own the science of our model development "feedback loop." While pre-training builds the base models, you define how we measure whether those models are safe, smart, and "Reddit-native," and you set the post-training methodology that turns base checkpoints into high-performing endpoints. You will define the Reddit Benchmark — our internal standard for rigorous model quality across both generation and representation — and own the evaluation science that the rest of the org's iteration depends on.

Responsibilities

Define the "Reddit Benchmark" evaluation standard: Own the methodology — not just the harness — for rigorously measuring model quality across Safety, Reasoning, representation/retrieval, and Reddit-specific knowledge. Decide what "Reddit-native" means in measurable terms and set the bar the org trains against.
Own evaluation reliability and statistical rigor: Establish the science behind trustworthy evals — judge variance, multi-sample scoring, inter-rater/inter-sample agreement, sampling and temperature effects, and calibration of automated judges. You are accountable for whether a benchmark delta is real or noise. Drive the practice of evaluation as a release gate — offline against frozen datasets, and pre-merge in CI/CD — so regressions are caught before endpoints ship.
Design model-as-a-judge methodology: Own judge selection, prompt design, calibration, and reliability for automated evaluation using frontier external models, enabling rapid, trustworthy iteration cycles.
Set post-training recipes and strategy: Design SFT recipes (data mixtures, curriculum, ablation strategy) that convert base models into helpful, well-aligned endpoints; partner with engineering to scale them.
Evaluate base and CPT checkpoints, not just endpoints: Design checkpoint-selection methodology across CPT experiments and LR studies, so we pick the right base before committing post-training compute.
Drive synthetic data generation strategy: Define and curate high-quality instruction and evaluation sets to improve generalization where human data is scarce.
Partner with Safety Engineering: Translate high-level safety policy into concrete classification metrics, probe sets, and CI/CD unit tests — including precision/recall at threshold, label-noise handling, and false-positive taxonomy for abuse detection (HHV).
Diagnose post-training instability: Dive into loss curves and eval logs to identify alignment tax and capability degradation, and recommend the fix.
Lead research direction: Set technical direction for evaluation and post-training across the team, mentor engineers and scientists, and represent the work internally (and externally where appropriate).

Required Qualifications

6+ years of professional ML experience (or PhD + 4+) with a direct focus on LLM post-training and evaluation.
PhD or MS in CS, ML, NLP, IR, or a related quantitative field — or equivalent industry research experience.
Deep expertise in evaluation reliability: judge/sample variance, multi-sample scoring, calibration, statistical significance, and the failure modes of automated evaluation.
Strong experience building custom, domain-specific evaluation harnesses (e.g., lm-eval-harness, Inspect AI, LightEval) — you know the strengths and limits of benchmarks like MMLU and GSM8K and when they don't apply, and you treat eval sets as versioned, frozen, regression-tracked code.
Experience evaluating both generation and representation/classification: model-as-a-judge for generative quality and precision/recall, PR-AUC, retrieval/MTEB-style metrics, gold-label denoising, and label-noise handling.
Deep understanding of Continuous Pre-training (CPT), Instruction Tuning (SFT), and how data quality shapes model behavior.
Fluency in Python; strong data-pipeline and eval-harness engineering (e.g., Hugging Face Transformers, vLLM, lm-eval-harness). Working knowledge of PyTorch and distributed training (FSDP2, DeepSpeed ZeRO-3) sufficient to direct and debug post-training runs.

Nice to Have

Experience with MLflow or similar experiment-tracking frameworks.
Familiarity with modern fine-tuning frameworks (Axolotl, TorchTune) and PyTorch-native training stacks (TorchTitan).
Synthetic data generation techniques (e.g., Self-Instruct).
Experience with preference optimization (DPO, RLHF, RLAIF, GRPO).
Publications in NLP/ML/FAccT or related venues, or other evidence of research leadership.
Experien

Responsibilities

Define the "Reddit Benchmark" evaluation standard: Own the methodology — not just the harness — for rigorously measuring model quality across Safety, Reasoning, representation/retrieval, and Reddit-specific knowledge. Decide what "Reddit-native" means in measurable terms and set the bar the org trains against.
Own evaluation reliability and statistical rigor: Establish the science behind trustworthy evals — judge variance, multi-sample scoring, inter-rater/inter-sample agreement, sampling and temperature effects, and calibration of automated judges. You are accountable for whether a benchmark delta is real or noise. Drive the practice of evaluation as a release gate — offline against frozen datasets, and pre-merge in CI/CD — so regressions are caught before endpoints ship.
Design model-as-a-judge methodology: Own judge selection, prompt design, calibration, and reliability for automated evaluation using frontier external models, enabling rapid, trustworthy iteration cycles.
Set post-training recipes and strategy: Design SFT recipes (data mixtures, curriculum, ablation strategy) that convert base models into helpful, well-aligned endpoints; partner with engineering to scale them.
Evaluate base and CPT checkpoints, not just endpoints: Design checkpoint-selection methodology across CPT experiments and LR studies, so we pick the right base before committing post-training compute.
Drive synthetic data generation strategy: Define and curate high-quality instruction and evaluation sets to improve generalization where human data is scarce.
Partner with Safety Engineering: Translate high-level safety policy into concrete classification metrics, probe sets, and CI/CD unit tests — including precision/recall at threshold, label-noise handling, and false-positive taxonomy for abuse detection (HHV).
Diagnose post-training instability: Dive into loss curves and eval logs to identify alignment tax and capability degradation, and recommend the fix.
Lead research direction: Set technical direction for evaluation and post-training across the team, mentor engineers and scientists, and represent the work internally (and externally where appropriate).

Required Qualifications

6+ years of professional ML experience (or PhD + 4+) with a direct focus on LLM post-training and evaluation.
PhD or MS in CS, ML, NLP, IR, or a related quantitative field — or equivalent industry research experience.
Deep expertise in evaluation reliability: judge/sample variance, multi-sample scoring, calibration, statistical significance, and the failure modes of automated evaluation.
Strong experience building custom, domain-specific evaluation harnesses (e.g., lm-eval-harness, Inspect AI, LightEval) — you know the strengths and limits of benchmarks like MMLU and GSM8K and when they don't apply, and you treat eval sets as versioned, frozen, regression-tracked code.
Experience evaluating both generation and representation/classification: model-as-a-judge for generative quality and precision/recall, PR-AUC, retrieval/MTEB-style metrics, gold-label denoising, and label-noise handling.
Deep understanding of Continuous Pre-training (CPT), Instruction Tuning (SFT), and how data quality shapes model behavior.
Fluency in Python; strong data-pipeline and eval-harness engineering (e.g., Hugging Face Transformers, vLLM, lm-eval-harness). Working knowledge of PyTorch and distributed training (FSDP2, DeepSpeed ZeRO-3) sufficient to direct and debug post-training runs.

Nice to Have

Experience with MLflow or similar experiment-tracking frameworks.
Familiarity with modern fine-tuning frameworks (Axolotl, TorchTune) and PyTorch-native training stacks (TorchTitan).
Synthetic data generation techniques (e.g., Self-Instruct).
Experience with preference optimization (DPO, RLHF, RLAIF, GRPO).
Publications in NLP/ML/FAccT or related venues, or other evidence of research leadership.
Experien