Training and validation sets of token-level routing labels for the R2R framework identify path-divergent tokens in reasoning tasks. These labels facilitate the training of a lightweight router that selectively switches between small and large language models to optimize inference efficiency.
Use Cases
- Train a lightweight neural router to predict when to switch from a small model to a large model using token-level routing labels
- Evaluate the efficiency of token-level routing strategies using the provided validation set
- Analyze path-divergent reasoning patterns by examining the distribution of routing labels across tokens
Strengths
- Contains token-level routing labels for small-large model orchestration
- Includes training and validation splits for the R2R (Roads to Rome) framework
- Identifies path-divergent tokens within reasoning sequences