Kiln's built-in interactive data generation system solves the time-consuming and laborious problem of traditional data annotation through a visual editing interface. The system supports users to quickly construct training samples through graphical drag-and-drop, and automatically generates structured data (JSON format) that meets the model training requirements. Its intelligent auxiliary function can automatically expand the relevant samples according to the keywords entered by the user, and provide real-time data quality verification tips.
The technical highlight of this feature is that it innovatively combines the rule engine and generative AI technology, which can ensure the accuracy of the generated data and also create diverse training samples through semantic expansion. The specific workflow includes: defining the data schema → setting the generation rules → previewing the generation results → batch exporting the dataset. The generated samples automatically contain complete annotations for multiple cue types (chain thinking/few samples/multi-samples).
Practical application cases show that the time to create 10,000 financial domain QA training data using the tool is reduced from 3 weeks to 4 hours for traditional manual labeling, and the data quality reaches the professional labeling level through automated calibration. This makes it easy for small and medium-sized teams to build high-quality domain-specific datasets.
This answer comes from the articleKiln: Simple LLM model fine-tuning and data synthesis tool, 0 code base to fine-tune your own small modelsThe
