Large Language Model Intern

Job Responsibilities

Train and optimize language models for Razer Software across the full pipeline—pretraining corpus construction, fine-tuning, evaluation, and deployment optimization.

You will work with senior data scientists and engineers on the model powering Razer Synapse’s device assistant, covering the full training pipeline:

• Corpus construction — cleaning, filtering, deduplication, and mixing of training data
• Fine-tuning — SFT, LoRA, and preference optimization runs
• Evaluation — quality, latency, and memory benchmarking against production constraints
• Deployment optimization — compression and quantization experiments for on-device targets

You will own discrete workstreams end-to-end (design → run → debug → analyze → iterate) rather than executing isolated tasks handed down by a mentor.

Learning Objectives

By the end of this internship, you will:

• Gain hands-on experience across the full LLM training lifecycle at production scale—not toy datasets
• Develop practical judgment in data engineering: how cleaning/filtering/mixing decisions propagate into model quality
• Run a rigorous ML experiment loop independently—hypothesize, train, evaluate, diagnose subtle regressions, and iterate
• See how deployment constraints (on-device latency and memory on Razer Synapse hardware) shape training and architecture decisions, including compression, quantization, and distillation tradeoffs
• Receive mentorship from senior engineers and visibility into how a production roadmap for a shipping AI feature is decided

Candidate Requirements

• Education: Current Bachelor’s, Master’s, or PhD student in Computer Science, AI, Data Science, or a related field
• Must-have knowledge: LLM training paradigms (pretraining, SFT, LoRA, preference optimization); Transformer/deep learning fundamentals; how deployment constraints (latency, memory) shape training decisions
• Must-have skills: Python; hands-on PyTorch model training; data engineering for training corpora (cleaning, filtering, deduplication, mixing); end-to-end experiment running (debugging, hyperparameter tuning, analysis); benchmarking quality/latency/memory
• Nice-to-have: Model compression (quantization, distillation), distributed training (multi-GPU/parallelism), Hugging Face/Accelerate/DeepSpeed, Linux environment
• Screening bar (hard requirement): Demonstrable hands-on model training/fine-tuning experience (coursework, research, internship, or open source). API-calling or prompt-engineering-only experience does not qualify.
• Strong positives: Trained a model from scratch (any scale), multi-GPU training, model compression/on-device deployment work, and top-tier publications (CVPR, NeurIPS, ICML, ACL, ICLR, EMNLP, etc.)

Pre-Requisites

Razer is proud to be an Equal Opportunity Employer. We believe that diverse teams drive better ideas, better products, and a stronger culture. We are committed to providing an inclusive, respectful, and fair workplace for every employee across all the countries we operate in. We do not discriminate on the basis of race, ethnicity, colour, nationality, ancestry, religion, age, sex, sexual orientation, gender identity or expression, disability, marital status, or any other characteristic protected under local laws. Where needed, we provide reasonable accommodations—including for disability or religious practices—to ensure every team member can perform and contribute at their best.

Are you game?