I'm a Machine Learning Engineer at Predibase, working on large language models and distributed training. I love working on open-source projects. 🚀
I maintain Ludwig, an open-source toolbox for low-code/no-code deep learning.
Some of my recent work includes:
- Using reinforcement fine-tuning to generate Triton kernels from PyTorch code snippets. Read More
- Co-creating Turbo LoRA, a parameter efficient training method to jointly fine-tune for quality and speculative decoding, improving throughput by upto 3.5x for fine-tuned LoRA adapters. Read More
- Developing a mixture of agents synthetic data generation algorithm for supervised instruction tuning that can beat K-shot GPT-4o using just 10 rows of data. Read More
- Co-authoring LoRA Land: 310 Fine-tuned LLMs that Rival GPT-4, A Technical Report. Read the paper.
- Speeding up LLM training by 15x. Read More
You can connect with me on LinkedIn to discuss all things AI 🤖.
If you like my work and want to support it, feel free to buy me a coffee ☕: https://buymeacoffee.com/arnavgarg