context-clip

Yes, context clipping truncates evaluation inputs and outputs to fixed context limits while preserving the original results. This allows you to generate context-variant evaluations for accurate comparison across different model cohorts.

To create context-limited evaluation variants, invoke the clipping tool with your cohort pattern, eval pattern, and desired context sizes. It parses cohorts, discovers matching evals, and generates multi-context variants automatically.

No, generating context-variant evals does not overwrite original results. The process tracks new eval IDs and updates evals.json while preserving the original evaluation data, allowing direct comparison between full and truncated contexts.

Use context clipping when you need to simulate smaller context limits across model cohorts. It helps evaluate how performance shifts when inputs and outputs are truncated, which is critical for deploying models to environments with restricted context windows.

The best way to truncate evaluation inputs for different context limits is using a CLI-driven workflow that parses cohorts and applies pattern matching. This automates multi-context variant creation and updates your evals.json file systematically.