Autonomous Data Selection with Language Models

• https://arxiv.org/pdf/2402.07625

• novel strategy that leverages base language models for autonomous data selection

  • meta-prompted language models as zero-shot verifiers

• Subsequent data selection leads to high downstream performance

• Assigns real-valued scores to training data

  • inspired by DPO
  • operates on the logits associated with ‘YES’ and ‘NO’ responses to meta-prompts
  • $\text{LM-Score}(\cdot )=\frac{exp(logit{(}^{\prime }YE{S}^{\prime }))}{exp(logit{(}^{\prime }YE{S}^{\prime }))+exp(logit{(}^{\prime }N{O}^{\prime }))}$

• The meta prompt can be composed of multiple questions

  • 
  • Thus, the complete score is defined as:
    • $\text{LM-Score}(Q_1,Q_2)=\text{LM-Score}(Q_1)\cdot \text{LM-Score}(Q_2)$

Math Data Selection

• Using Qwen-72B model
•