CRAFT: LLM Capability Diagnosis
Diagnoses LLM weaknesses by clustering rubric criteria into a hierarchy, scoring performance at each level, and identifying specific failure points for targeted fine-tuning data generation.
When to use
When you have a rubric-based evaluation dataset and need to understand not just where your model fails, but why, in order to generate targeted data for fine-tuning and improve specific capabilities.
How to apply
Extract capability descriptions from prompt-rubric pairs in your evaluation dataset.
Cluster these descriptions into a hierarchical tree structure.
Score the target LLM's performance at each node within the hierarchy.
Identify nodes where the model performs poorly.
Use these low-performing nodes to direct the generation of supervised fine-tuning data.
Glossary
- Capability Descriptions
- Textual explanations of specific skills or behaviors an LLM is expected to exhibit, derived from evaluation prompts and their associated scoring rubrics.
- Hierarchical Tree Structure
- An organized arrangement of data where items are linked in a parent-child relationship, forming a tree-like structure that represents levels of abstraction or specificity.
- Performance Scoring
- The process of assigning a quantitative score to an LLM's output based on predefined criteria, indicating how well it meets the requirements of a given task or capability.
- Supervised Fine-Tuning Data
- A dataset consisting of input-output pairs used to further train a pre-trained LLM, where the outputs are considered correct or desirable for specific tasks.
- Prompt-Rubric Pairs
- Sets of data that include a specific instruction or question given to an LLM (the prompt) and a detailed set of criteria used to evaluate the LLM's response (the rubric).