top of page
pmmucsd_a_field_of_small_As_purple_geometric_shapes_that_seem_ab322f7d-61e1-4d60-b807-97c0

Introducing AutoTune: Automated Fine-Tuning Data Generation

Oct 3, 2024


Fine-tuning Dataset for Customer Service Agents using AutoTune

We're excited to announce AutoTune, a powerful new tool to automatically generate high-quality training data for fine-tuning language models. With AutoTune, you can go from an initial prompt to a downloadable fine-tuning dataset in just a few clicks.


How does it work? 


The process is remarkably simple:

  1. Sign into Aligned’s platform to access our data creation tools. It's free and easy - no strings attached.



Login to Aligned's Data Platform
Create a new AutoTune Dataset

  1. Create a new AutoTune project and enter your initial prompt. This prompt represents an ideal system message to prime the model for your use case.


Add your initial prompt
  1. Add seed data (ideal questions and responses) if you have it. AutoTune generates an initial batch of diverse questions and example responses based on your prompt and seed data.

Add seed data if you have it
  1. Provide feedback on the generated questions and responses. Was the data on target or does it need adjustment? Your feedback is used to automatically rewrite and improve the prompt.


    Review the initial output and give feedback

    Review and modify the revised prompt


  2. AutoTune conducts additional rounds of data generation and evaluation, progressively refining the prompt and response quality based on comparisons between a state-of-the-art model and a smaller reference model.


View the LLM-based judging of output
  1. The top performing prompt variations and responses are used as "few-shot" examples to generate a final fine-tuning dataset of 100+ question and response pairs. Simply review the data, pick the responses you like best, and download the data in JSON Lines format, ready to use with popular fine-tuning platforms.


Review the final prompt and examples before creating the data

Review the generated dataset and remove bad samples

Access, download and edit your datasets

Under the hood, AutoTune leverages the power of several state of the art models, to analyze your feedback and synthesize data at each step in the process. By systematically exploring variations of your prompt and evaluating the outputs, AutoTune zeros in on the most effective phrasing and style to elicit high-quality responses tailored to your application. 


Potential use cases are endless.

  • Fine-tune an AI assistant for your specific knowledge domain, ensuring it provides accurate, relevant information to user queries

  • Train a code generation model to follow your project's coding conventions and architectural patterns

  • Customize a grammar and style checking model for your organization's preferred voice and terminology

  • Create a model that role plays different characters or personas for gaming and entertainment

Best of all, AutoTune significantly reduces the labor in creating and curating data. The complex work of prompt engineering and data curation is fully automated, allowing developers to focus on building great applications with fine-tuned models.


We can't wait to see what you create with AutoTune! The tool is now available in limited beta, just sign into Aligned’s data platform to get started.

Happy fine-tuning!

83 views0 comments

Recent Posts

See All

Fine-Tuning Best Practices

Every large language model runs on a specific set of parameters or weights that determine how the model behaves.  Cutting-edge LLMs end...

Comments


bottom of page