Refuel Cloud: Platform to label, clean and enrich your data at scale using LLMs

Schedule a Demo

January 21, 2024

Refuel Team
Refuel Team

We started Refuel with the mission to accelerate the era of AI abundance, and we wanted to address the first, and most important challenge – the availability of clean, labeled data. As they say: “Garbage in, garbage out” – and having spoken to hundreds of teams building data-driven software products and AI models, rarely do the teams have the data they need for the products they desperately want to build. 

That barrier changes today – we’re thrilled to announce Refuel Cloud, a radical new way for businesses to get the datasets they need, whenever they need and at any scale. It’s designed for teams working on a wide range of problems, such as classification, data extraction, entity resolution and recognition, and is being used by companies in many different industries. What’s more – Refuel Cloud has been used to produce 10B+ annotations already across our early partners and customers. 

You don’t need to be an ML expert to use Refuel Cloud, and you can get started in minutes. Refuel Cloud is in closed beta - sign up here.

Why Refuel Cloud? 

Until recently, the only way for teams to clean, label and enrich datasets was through manual effort. You could choose to spend weeks of your team’s time getting your data in order, or work with some third party that doesn’t really understand your domain – either way, it’s slow, painful and with no certainty of data quality. Alternatively, you can choose to use LLMs directly, but the accuracy is not high enough when using an LLM off-the-shelf – even GPT-4 falls short for most business needs. 

Refuel Cloud aspires to be the platform that gives you the data quality of a human expert, with the speed and scale that a machine can deliver. Working with early partners, we have found Refuel Cloud to help teams in three important ways when using LLMs for data labeling, cleaning and enrichment. 

For high accuracy, provide feedback + fine-tune, don’t prompt engineer 🛠️

Refuel makes it easy to ascend this ladder of complexity (adapted from Andrej Karpathy’s tweet)

Refuel helps teams improve performance on any task very quickly, and it involves the following three components:

  1. Task guidelines: Users can provide guidelines for how they want the LLM to clean, label or enrich their data in simple natural language – imagine just talking to another human.
  2. Examples and feedback for few-shot learning: Refuel’s simple interface makes it easy for a user to provide feedback on any data labeled by the LLM, and Refuel continues to surface the examples that are hardest for the LLM at any given time. These examples are then used as helpful examples when labeling the next batch of data. 
  3. Model fine-tuning: Fine-tuning LLMs is what consistently takes you to 99%+ accuracy on every task, but setting up the right infrastructure, hyperparameter tuning and dataset selection is challenging. Refuel Cloud hides away the complexity and makes it a seamless experience, so you can systematically improve accuracy for every task. 

If your team is spending weeks just tuning prompts, and not seeing any meaningful gains in performance, you’re likely stuck in prompt engineering hell. Instead, spend that time capturing feedback on difficult examples within Refuel and just customize an LLM directly. 

Compared to human data labeling – Refuel Cloud is 100x faster and better quality. 

“We compared Refuel against our existing manual data labeling solution, and not only did Refuel’s LLM produce more accurate labels as validated by human teams, but their speed for large datasets was astounding – 1 day compared to 5 weeks”, says Charles Zhu, VP of Product Management at Enigma 

LLMs 🤝 Humans > LLMs or Humans

There are many real world applications where mistakes are extremely costly. Imagine, extracting data from documents for a loan application, or parsing through high-value insurance claims documents – a single mistake can cause reputational damage, or $$$$ damage. 

These applications have traditionally required large teams doing tons of manual work. Even in the age of LLMs, teams are unsure whether to try and automate these processes, or to have a human-in-the-loop consistently. 

Refuel Cloud provides a better way – our research has demonstrated reliable confidence estimation for LLM outputs that makes it possible for teams to intelligently route low-confidence examples for manual review, while 90% of the process can be fully automated. With every piece of human feedback, the performance of Refuel on that task continues to improve. 

Balance accuracy and automation rate with reliable, calibrated confidence scores from Refuel's LLMs

Scaling 📈 and data operations at low-latency ⚡

Teams are often comfortable working with LLMs during development-time on small datasets, but the costs for production traffic and the high latency from the largest models makes many applications infeasible to run in production environments. Refuel makes it a one-click experience to deploy a labeling application to run in real-time, and at any scale customers want. 

Since Refuel Cloud comes equipped with Refuel LLM – an extremely capable LLM for data labeling – we make it easy to scale data volumes seamlessly and without having to be rate limited by external LLMs provided through APIs. 

“Beni has a product catalog of 200M+ items and ensuring clean, structured data is an ongoing challenge. Using Refuel’s customized LLMs, we were able to label millions of items and improve accuracy on a key attribute from 46% to 87%”, says Celine Lightfoot, CTO of Beni. “What would have taken us months, only took a few days with Refuel.”

Our early partners have used Refuel to deliver billions of data annotations already – and this number is growing quickly. 

Get started

No matter whether you’re extracting information from PDF documents, or tagging financial transactions, or detecting abusive content on your platform, Refuel Cloud can help. We can’t wait to see what you’ll build! 

Sign up here to get started with Refuel Cloud: