A cross-task collaborative thought chain distillation method, device, system, and storage medium

Abstract

The present invention discloses a cross-task collaborative thought chain distillation method, device, system, and storage medium, comprising: step S1, generating thought chains for a reasoning task dataset; step S2, ranking the difficulty of the tasks to be trained based on the thought chains; step S3, orthogonally grouping and training student models based on the thought chains and task difficulty; and step S4, distilling the cross-task collaborative thought chains based on the LoRA expert combination. The technical solution of the present invention prevents negative transfer by implicitly isolating parameters between unrelated tasks, further enhancing the reasoning capability of small models.

Claims

1. A cross-task collaborative thought chain distillation method, characterized by: include: Step S1: Generate a thought chain for the reasoning task dataset; Step S2: sorting the difficulty of the tasks to be trained according to the thought chain; Step S3: conducting orthogonal group training on student models according to thinking chains and task difficulty; Step S4: Distill cross-task collaborative thinking chains based on the LoRA expert combination; Each reasoning task contains complex reasoning questions and the correct answer Where t≤m; using Zero-shot-CoT, that is, adding Let's think step by step after the reasoning problem. Guide the LLM with CoT capability to solve the problem Where i≤n, generating multi-stage explanations and predict the results According to the correct answer and prediction results Choosing the right multi-stage explanation use Form a sample The training method of step S3 is LoRA fine-tuning, and the LoRA fine-tuning strategy includes similarity detection and group orthogonal regularization; among them, similarity detection is used to detect whether there are tasks similar to the current task in the trained tasks before training the current task; group orthogonal regularization is added between the LoRA experts of dissimilar tasks to constrain the current task to update parameters in a direction orthogonal to the gradient of the dissimilar tasks.

2. A cross-task collaborative thinking chain distillation device for implementing the cross-task collaborative thinking chain distillation method according to claim 1, characterized in that: include: The first processing module is used to generate a thought chain for the reasoning task dataset; The second processing module uses the thought chain to sort the difficulty of the tasks being trained; The third processing module uses orthogonal grouping training of student models based on thinking chains and task difficulty; The fourth processing module uses cross-task collaborative thought chain distillation based on LoRA expert combination.

3. A cross-task collaborative thought chain distillation system, characterized by: include: A memory and a processor, wherein the memory stores a computer program run by the processor, and when the computer program is run by the processor, the cross-task collaborative thinking chain distillation method as described in claim 1 is executed.

4. A storage medium, characterized in that The storage medium stores a computer program, which executes the cross-task collaborative thinking chain distillation method as described in claim 1 when running.

Images