24 lines (14 loc) · 776 Bytes

CHANGELOG

Inspired from Keep a Changelog

[Unreleased]

Breaking Changes

Features

Introduced dynamic percentile-based relevance thresholding for binary-dependent metrics (Precision, MAP) to replace hard-coded j > 0 mapping (#394)

Enhancements

Bug Fixes

Fixed thread pool starvation in LLM judgment processing (#387)

Infrastructure

Documentation

Maintenance

Refactoring

Extract reusable BatchedAsyncExecutor; migrate LlmJudgmentTaskManager and ExperimentTaskManager to use it (#392)