AnonymousORG01/QASM-Eval

QASM-Eval 是一个基于 OpenQASM 3 的数据集，用于在硬件相关的量子编程任务上训练和评估语言模型。它专注于近期量子工作流程中至关重要的能力，包括时序感知编程、经典控制、脉冲级操作以及将这些元素结合在单个程序中的混合工作流。每个示例都是一个 OpenQASM 3 代码补全任务：提示包含一个完整的 QASM 程序，其中一个缺失的核心区域用 TODO 风格的自然语言描述标记，模型需要仅生成标记之间的缺失代码块。数据集覆盖四个领域：时序（延迟、持续时间和调度敏感行为）、经典（测量、寄存器、条件语句和经典控制）、脉冲（帧、波形、defcal 和脉冲级校准逻辑）以及复杂（结合量子、时序、脉冲和经典行为的混合任务）。数据集分为训练集（4000 条记录，每个领域 1000 条）和测试集（100 条记录，每个领域 25 条），提供 JSONL 和 Parquet 格式。每个记录包含任务ID、领域、提示、补全内容、规范解决方案和测试片段等字段。

QASM-Eval is an OpenQASM 3 dataset for training and evaluating language models on hardware-facing quantum programming tasks. It focuses on capabilities that matter in near-term quantum workflows, including timing-aware programming, classical control, pulse-level operations, and mixed workflows that combine these elements in a single program. Each example is an OpenQASM 3 code-completion task: the prompt contains a full QASM program with one missing core region marked by a TODO-style natural-language description, and the model is expected to generate only the missing block between the markers. The dataset covers four domains: timing (delays, durations, and schedule-sensitive behavior), classical (measurement, registers, conditionals, and classical control), pulse (frames, waveforms, defcal, and pulse-level calibration logic), and complex (mixed tasks combining quantum, timing, pulse, and classical behavior). It is split into train (4,000 records, 1,000 per domain) and test (100 records, 25 per domain) sets, available in JSONL and Parquet formats. Each record includes fields such as task_id, domain, prompt, completion, canonical_solution, and test.