feat: wire --inflate flag into eval CLI for 400K token scenario testing#386
Merged
Conversation
Adds --inflate <tokens> flag to eval run.ts that inflates scenarios before running them. Uses inflateScenario() from inflate.ts. First results at 400K tokens (PR-2 implicit preferences): Lore: 4.20 Tail-window: 2.90 At realistic conversation lengths, Lore decisively outperforms tail-window on 6/8 questions.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds
--inflate <tokens>flag to the eval CLI that inflates scenarios to a target token count before running them. This enables fair baseline comparison at realistic conversation lengths.First Results at 400K Tokens
PR-2 (implicit preferences):
Lore decisively outperforms tail-window at realistic conversation lengths. At 400K tokens, tail-window can only fit the last 80K tokens (dropping preferences stated early in the conversation), while Lore's distillation preserves them.
Lore wins 6/8 questions. Tail-window only wins on 2 questions where the relevant facts happen to be in its surviving 80K window.
Usage
Files Changed
packages/core/eval/run.ts—--inflatearg parsing + loggingpackages/core/eval/types.ts—inflateTokensfield on EvalConfigpackages/core/eval/harness.ts— inflation before scenario execution