Data screening method, device and equipment and computer storage medium
By using consensus metrics and similar strategy metrics to filter out high-quality training data during the language model fine-tuning process, the problem of selecting valuable training data from massive datasets is solved, thus improving the training quality and effectiveness of the model.
CN121524635BActive Publication Date: 2026-06-12TAOBAO CHINA SOFTWARE
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TAOBAO CHINA SOFTWARE
- Filing Date
- 2026-01-13
- Publication Date
- 2026-06-12
AI Technical Summary
Technical Problem
In the process of fine-tuning a language model, how can we efficiently and cost-effectively filter out valuable training data from massive and noisy data sources to improve the training quality and effectiveness of the model?
Method used
By acquiring multiple preference data samples, we determine their consensus index values and policy index values under different preference dimension parameters, and then select target training samples corresponding to the model fine-tuning task.
Benefits of technology
This enables efficient and low-cost screening of training data that is valuable for model training operations, thereby improving the training quality and effectiveness of language models.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN121524635B_ABST
Abstract
Embodiments of the present application provide a data screening method, device and equipment and a computer storage medium; wherein the data screening method comprises: obtaining a plurality of preference data samples, the preference data sample comprising input information, positive output samples and negative output samples under at least one preference dimension parameter; determining a consensus index value of the plurality of preference data samples under different preference dimension parameters respectively, the preference data sample corresponding to an original preference dimension parameter, the consensus index value being used to identify the recognition degree of the preference data sample under a non-original preference dimension parameter; determining a same strategy index value corresponding to the plurality of preference data samples respectively, the same strategy index value being used to identify the adaptation degree between the preference data sample and a language model to be trained; and screening target training samples corresponding to a model fine-tuning task from the plurality of preference data samples based on the consensus index value and the same strategy index value corresponding to the plurality of preference data samples respectively.
Need to check novelty before this filing date? Find Prior Art