Senior AI Model Training Specialist — AdamW Benchmarking

Fractional consulting role — remoteDevOpt Labs is looking for a senior practitioner with deep hands-on experience training and tuning models with AdamW across the full model-training lifecycle. It is imperative that the candidate understand and have extensive practical experience with:pretraining, continued pretraining (CPT), fine-tuning, post-training, RL/RLHF-style training, and other downstream adaptation regimes. The primary job is to train a variety of model types and sizes using AdamW and determine the best achievable AdamW settings for each workload and training stage. These results will be used in publication-quality comparisons against AERO, our proprietary training technology, which works very differently from AdamW. The AdamW baseline must be strong enough to withstand rigorous scientific scrutiny. Our methodology is guided by the September 2025 Stanford paper “Fantastic Pretraining Optimizers and Where to Find Them.” The goal is not to reduce rigor. It is to find the most cost-efficient path to the true AdamW optimum, using intelligent sweep design, prior results, scaling behavior, and expert judgment rather than unnecessary brute-force experimentation. ResponsibilitiesTrain models using AdamW across pretraining, CPT, fine-tuning, post-training, RL, and related regimes.Find the best AdamW configuration for each model, scale, dataset, and training stage.Tune learning rate, warmup, weight decay, betas, batch size, and related parameters as needed.Design efficient sweep strategies that minimize GPU cost without compromising the quality of the optimum found.Determine when poor candidates can be eliminated early and when full endpoint runs are required.Produce reproducible, publication-quality AdamW baselines.Help ensure AERO-vs-AdamW comparisons withstand expert scientific review. Required experienceCandidates should have personally operated substantial large-scale training runs and have strong experience with:AdamW optimizationBillion +-parameter modelsPretraining and continued pretrainingFine-tuning and post-trainingRL/RLHF or comparable post-training methodsMulti-GPU H100/A100 trainingHyperparameter sweepsValidation-loss analysisMulti-seed benchmarkingReproducible scientific experiments This is not a programming role. We need someone who deeply understands how training behavior, optimizer settings, schedules, and optimal configurations change across the entire lifecycle of a model. Experience with other popular optimizers will be useful. CompensationApproximately $175–$300/hour, depending on experience. To applyPlease briefly describe:The largest models you have personally trained.Your experience across pretraining, CPT, fine-tuning, post-training, and RL.Your experience tuning AdamW.How you would find the true AdamW optimum while minimizing unnecessary full training runs.Your hourly rate and availability.