October 2023 Summaries

Refuel LLM, a large language model specifically designed for data labeling and enrichment tasks, has been launched, demonstrating superior performance over trained human annotators and several other language models, including GPT-3.5-turbo and PaLM-2, across a benchmark of 15 text labeling datasets. Built on the Llama-v2-13b base model, Refuel LLM is trained on over 2500 unique datasets spanning various categories and can be fine-tuned on target domains to further enhance performance and reduce inference costs. The model is publicly accessible via the LLM labeling playground and the Autolabel open-source library, with plans for a more detailed technical summary and open-source release in the future. Fine-tuning experiments have shown that Refuel LLM can achieve superhuman performance quickly, outperforming GPT-4 in certain tasks with minimal training time. The model's training encompassed diverse instructions to ensure accurate label output without parsing, and it is now available for use through various platforms, with further details and support available via Refuel Cloud.

Refuel Cloud is an innovative platform designed to address the challenge of obtaining clean, labeled data for AI models by providing businesses with the datasets they need at any scale and with high accuracy. It empowers teams working on tasks such as classification and data extraction by combining human-level data quality with the speed and scalability of machine learning. Refuel Cloud simplifies the process by allowing users to provide task guidelines in natural language, offer feedback for few-shot learning, and fine-tune large language models (LLMs) to achieve over 99% accuracy. Unlike traditional manual data labeling, Refuel Cloud is significantly faster and more efficient, reducing the time required from weeks to days. Additionally, it features reliable confidence estimation to intelligently route low-confidence examples for manual review, enabling a balance between accuracy and automation. With its ability to handle large data volumes without being constrained by external LLMs, Refuel Cloud is already being used across industries, helping companies like Beni and Enigma achieve substantial improvements in data labeling accuracy and speed.