Model Experimentation Made Easier

Overview

Refuel.AI built Refuel LLM, a purpose-built model for data labeling and enrichment, by instruction-tuning a Llama-v2-13b base model on more than 5 billion tokens using Databricks Training infrastructure. The initial training run produced a 78% increase in label quality, and subsequent fine-tuning on a cluster of 8x H100s added a further 16% performance gain, outperforming trained human annotators and several other LLMs on a 15-dataset text labeling benchmark.

The challenge

Data labeling requires considerable resources and time; manual labeling is costly and vulnerable to human errors, and human-in-the-loop methods only marginally speed up the process and increase accuracy.

The solution

Refuel.AI built Refuel LLM, a purpose-built model for data labeling and enrichment tasks, instruction-tuned on more than 5 billion tokens across more than 2,500 unique tasks on top of a Llama-v2-13b base model. The team trained close to 50 models over almost three months on Databricks Training infrastructure, with initial training runs of about three days each, then fine-tuned the model on a cluster of 8x H100s within Databricks Training to further improve performance and reduce prompt lengths.

Reported business value

Refuel LLM outperformed trained human annotators and other leading LLMs (GPT-3.5-turbo, PaLM-2, Claude) across a benchmark of 15 text labeling data sets, and the initial release attracted over ten thousand users accessing the Refuel LLM cloud or playground.