LongForge: Improving long-context instruction data via evidence-assertion reasoning graphs
In brief
- LongForge measures the quality of long-context instruction data using reasoning graphs.
- The tool uncovers major flaws: low breadth and depth of reasoning in existing datasets.
- Models trained with LongForge outperform baselines by 2.2 to 4.7 points across six benchmarks.
LongForge introduces a method to evaluate and generate long-context instruction data by analyzing their reasoning structure. The tool identifies major flaws in existing datasets and improves model performance on six benchmarks.
This summary gives you the essentials. The full article, its figures and quotes are on cs.AI updates on arXiv.org.
Read the original articleAI-generated summary of the source, checked by rules. Report an error