Action Data for Large Action Model (LAM)

Overview

SoftAge generated high-quality action data to train Rabbit’s action models.

Rabbit aimed to develop action models for consumer applications such as Uber and DoorDash, where action execution data and prompts were needed. SoftAge partnered with Rabbit to deliver a dataset tailored to the unique demand for consumer app-focused action data.

Challenge

Creating action prompts required a delicate balance between diversity and practicality. While diversity fosters model adaptability, prompts also needed to reflect commonly performed actions to ensure real-world relevance. Furthermore, Rabbit's success hinged on achieving 100% annotation accuracy for the data to meet the high standards of consumer applications.

Solution

SoftAge implemented a comprehensive approach to address Rabbit’s needs:

  • Balanced Action Prompt Creation: A dedicated team meticulously designed prompts that struck the right balance between diverse and frequently executed actions. This ensured the model could handle a wide range of scenarios while excelling in commonly performed tasks.
  • User Familiarization with Tools: To maximize accuracy and consistency, SoftAge provided users with dedicated time and training to become proficient with the tools for which action data was needed. This measure improved productivity and reduced errors, leading to higher-quality data.
  • Rigorous Quality Assurance: Each dataset underwent thorough quality checks by experienced annotators to guarantee 100% annotation accuracy. Our multi-step review process ensured that Rabbit received error-free and contextually appropriate action data.

Result

SoftAge successfully delivered a high-quality, diverse dataset of action prompts spanning key consumer app functionalities. Through this collaboration, SoftAge reinforced its expertise in generating action data for AI models.