hle-benchmark-evolver
由 wanng-ide 提供
Browser & Automation简介
Runs HLE-oriented benchmark reward ingestion and curriculum generation for capability-evolver.
描述
该技能的功能
Ingests question-level HLE benchmark results into the capability-evolver reward system and generates curriculum signals for evolution cycles. Converts per-question accuracy data into prioritized queues that focus subsequent training on easier questions first.
Closes the loop between benchmark evaluation and capability evolution by automating reward ingestion, curriculum prioritization, and solidification in a single pipeline command.
何时使用
- Processing an HLE report JSON into reward and curriculum signals
- Running a multi-cycle evolution pipeline with automatic re-evaluation
- Checking current benchmark accuracy and curriculum stage snapshot
- Targeting a specific HLE score threshold using easy-first question queues
- Identifying which subjects and modalities to focus on next
示例工作流
User asks: ingest my HLE report and generate curriculum signals
- 1
Validate the report JSON exists at the provided path
- 2
Ingest report into capability-evolver benchmark reward state
- 3
Generate curriculum signals: benchmark_*, curriculum_stage:*, focus_subject:*, question_focus:*
- 4
Print compact result summary as JSON
JSON with benchmark_id, accuracy, reward, trend, curriculum_stage, queue_size, focus_subjects, and next_questions
安装
clawhub install wanng-ide/hle-benchmark-evolver运行 CLI 命令,或把这段提示词粘贴到 Claude Code、Cursor、Codex 或你在用的任意编程智能体里。