# QASM-Eval 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. This Hugging Face dataset release contains the benchmark data itself, a lightweight evaluation entry point, and the scripts needed for local checking: - `data/` - `scripts/` - `run_test.py` For the full project repository, documentation, and broader codebase, see the GitHub repo: [AnonymousUser-01/QASM-Eval](https://github.com/AnonymousUser-01/QASM-Eval). Two models fine-tuned on QASM-Eval are also available on Hugging Face: - [Llama-3.1-8B-Instruct-QASM-Eval](https://huggingface.co/AnonymousORG01/Llama-3.1-8B-Instruct-QASM-Eval) - [Llama-3.3-70B-Instruct-QASM-Eval](https://huggingface.co/AnonymousORG01/Llama-3.3-70B-Instruct-QASM-Eval) ## Dataset Summary 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 description. 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 - `complex`: mixed tasks combining quantum, timing, pulse, and classical behavior ## Splits | split | records | composition | | --- | ---: | --- | | `train` | 4,000 | 1,000 timing, 1,000 classical, 1,000 pulse, 1,000 complex | | `test` | 100 | 25 timing, 25 classical, 25 pulse, 25 complex | Files are provided in both JSONL and Parquet format: - `data/train.jsonl` - `data/test.jsonl` - `data/train.parquet` - `data/test.parquet` ## Data Fields Each record contains the following fields: | field | description | | --- | --- | | `task_id` | Stable task identifier | | `domain` | One of `timing`, `classical`, `pulse`, or `complex` | | `prompt` | OpenQASM 3 program with the target core block replaced by a natural-language TODO | | `completion` | Reference answer for the missing core block only | | `canonical_solution` | Full reference OpenQASM 3 program | | `test` | Minimal evaluation snippet showing how the completion is checked | The central prompt structure looks like: ```qasm // === CORE_TASK_START === // TODO(core task): natural-language description of the missing QASM behavior // === CORE_TASK_END === ``` Models should output only the code that belongs between these two markers. ## Quick Start: ### 1. Evaluate an LLM on QASM-Eval Use `run_test.py` for end-to-end model benchmarking on the released test split. The script is configured directly at the top of the file rather than through command-line arguments. Important configuration fields include: - `RUN_MODE`: choose `"random_sample"` or `"all"` - `NUM_RANDOM_SAMPLES`: total number of randomly selected test tasks when using random mode - `PASS_AT_K`: number of model samples generated per task, and the maximum `k` reported in pass@k - `LLM_CHOICE`: model name - `LLM_PROVIDER`: backend provider such as `nebius`, `openai`, `nscale`, or `huggingface` - `LLM_KEY`: API key for the chosen provider - `LLM_TEMPERATURE`: generation temperature - `LLM_MAX_TOKENS`: max response length The script reads `data/test.jsonl`, queries the model repeatedly for each selected task, evaluates every generated answer, and writes run artifacts into `test_runs/`. To launch an evaluation run: ```bash python3 run_test.py ``` The output includes a detailed report with the overall pass@k metric and pass rates for each task category. Per-sample error types is also available, the scripts in ./scripts contain the full simulator implementations and the associated evaluation logic. ### 2. Use QASM-Eval as a training dataset Load the released JSONL or Parquet files directly for supervised fine-tuning, analysis, or prompt construction. Typical field usage: - `prompt`: model input - `completion`: target core block only - `canonical_solution`: full reference program You can also load the Parquet files if that better fits your training stack. ## How To Load Separately Using the Hugging Face `datasets` library: ```python from datasets import load_dataset dataset = load_dataset("AnonymousORG01/QASM-Eval") train_data = dataset["train"] test_data = dataset["test"] print(train_data[0].keys()) print(train_data[0]["domain"]) print(train_data[0]["prompt"][:300]) ``` If you prefer to work with the local files directly: ```python import json with open("data/train.jsonl", "r", encoding="utf-8") as f: first_record = json.loads(next(f)) print(first_record["task_id"]) print(first_record["completion"]) ``` ## Citation This project is currently maintained in anonymous form for review purposes. A finalized citation will be added after the review process.