Description
Problem
The way editing works is non-optimal. Oftentimes edits fail and have to be repeated. And apparently the way the edit command is implemented can significantly impact model performance.
Solution
omp investigated this question and implemented something called hashline. Edits are now more likely to succeed. This reduces tokens used, saves $$, and apparently even improves model performance.
(FYI: model performance is improved anytime tokens are saved.)
Would be cool if Crush adopted this approach too!
Description
Problem
The way editing works is non-optimal. Oftentimes edits fail and have to be repeated. And apparently the way the edit command is implemented can significantly impact model performance.
Solution
omp investigated this question and implemented something called hashline. Edits are now more likely to succeed. This reduces tokens used, saves $$, and apparently even improves model performance.
(FYI: model performance is improved anytime tokens are saved.)
Would be cool if Crush adopted this approach too!