18 Nov
|
Opkey
|
Bengaluru
Job Description
We are hiring an engineer to fine-tune, train and evaluate large language models (LLMs) and vision-language models (VLMs) for production use. You will own models end to end: data, training, evaluation. The role suits someone who solves hard problems from first principles and understands why techniques work, not just how to run them.
Key responsibilities
• Fine-tune open-weight LLMs and VLMs for domain-specific tasks.
• Apply full fine-tuning and parameter-productive methods (LoRA, QLoRA, adapters).
• Run post-training and alignment: supervised fine-tuning (SFT), instruction tuning, and preference optimization (DPO, RLHF/PPO, GRPO).
• Build data pipelines: collect, clean, deduplicate, filter and synthesize text and image text training data.
• Design evaluation: task benchmarks, LLM-as-judge setups, human eval, and regression suites that catch quality drops.
• Train efficiently on multi-GPU setups using distributed training (DeepSpeed, FSDP),
mixed precision and gradient checkpointing.
• Debug training failures such as loss spikes, divergence, overfitting, catastrophic forgetting and hallucination, and find the root cause.
• Optimize models for inference through quantization (GPTQ, AWQ, GGUF), distillation and serving with vLLM or TGI.
• Read and reproduce recent research papers, and turn promising ideas into working experiments.
• Track experiments rigorously and share findings clearly with engineering and product teams. Required
qualifications:
• 2–3 years in machine learning or deep learning.
• Shipped at least one fine-tuned model to production or a real user-facing use case.
• Strong AI fundamentals and Problem solving.
• Bachelor's or Master's in Computer Science, AI, Mathematics or a related field (or equivalent hands-on work)
📌 AI Engineer (Bengaluru)
🏢 Opkey
📍 Bengaluru