Staff Research Engineer, Post-training & Evaluation
Reddit · Website ↗
Job Description
- 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).
Requirements
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.
- Experience evaluating multimodal models (embeddings, hateful-memes-style classification).