isage-data 0.2.2.0

SAGE Data - Unified data loaders for memory benchmark datasets (LongMemEval, Locomo, MemAgentBench, etc.)

SAGE Data ��

Dataset management module for SAGE benchmark suite

Provides unified access to multiple datasets through a two-layer architecture:

• Sources: Physical datasets (qa_base, bbh, mmlu, gpqa, locomo, orca_dpo)
• Usages: Logical views for experiments (rag, libamm, neuromem, agent_eval)

Quick Start

./quickstart.sh
source .venv/bin/activate

Or manual steps:

from sage.data import DataManager

manager = DataManager.get_instance()

# Access datasets by logical usage profile
rag = manager.get_by_usage("rag")
qa_loader = rag.load("qa_base")  # already instantiated
queries = qa_loader.load_queries()

# Or fetch a specific data source directly
bbh_loader = manager.get_by_source("bbh")
tasks = bbh_loader.get_task_names()

🛠️ CLI 使用方式（精简版）

安装后可直接使用 sage-data 命令：

sage-data list               # 显示数据源状态（已下载/缺失/远程）
sage-data usage rag          # 查看某个 usage 的数据映射
sage-data download locomo    # 下载指定数据源（仅支持部分源）

# 选项
sage-data list --json        # JSON 输出，便于脚本处理
sage-data --data-root /path  # 指定自定义数据根目录

当前支持自动下载的源：locomo, longmemeval, memagentbench, mmlu。 其他如 gpqa, orca_dpo 采用按需在线加载（Hugging Face），qa_base/bbh 等随包内置。

Available Datasets

Dataset	Description	Download Required	Storage
qa_base	Question-Answering with knowledge base	❌ No (included)	Local files
locomo	Long-context memory benchmark	✅ Yes (python -m locomo.download)	Local files (2.68MB)
bbh	BIG-Bench Hard reasoning tasks	❌ No (included)	Local JSON files
mmlu	Massive Multitask Language Understanding	📥 Optional (python -m mmlu.download --all-subjects)	On-demand or Local (~160MB)
gpqa	Graduate-Level Question Answering	✅ Auto (Hugging Face)	On-demand (~5MB cached)
orca_dpo	Preference pairs for alignment/DPO	✅ Auto (Hugging Face)	On-demand (varies)

See examples/ for detailed usage examples.

📖 Examples

python examples/qa_examples.py            # QA dataset usage
python examples/locomo_examples.py        # LoCoMo dataset usage
python examples/bbh_examples.py           # BBH dataset usage
python examples/mmlu_examples.py          # MMLU dataset usage
python examples/gpqa_examples.py          # GPQA dataset usage
python examples/orca_dpo_examples.py      # Orca DPO dataset usage
python examples/integration_example.py    # Cross-dataset integration

License

MIT License - see LICENSE file.

🔗 Links

• Repository: https://github.com/intellistream/sageData
• Issues: https://github.com/intellistream/sageData/issues

❓ Common Issues

Q: Where's the LoCoMo data?
A: Run python -m locomo.download to download it (2.68MB from Hugging Face).

Q: How to download MMLU for offline use?
A: Run python -m mmlu.download --all-subjects to download all subjects (~160MB).

Q: GPQA access error?
A: You need to accept the dataset terms on Hugging Face: https://huggingface.co/datasets/Idavidrein/gpqa

Q: How to use Orca DPO for alignment research?
A: Use DataManager.get_by_source("orca_dpo") to get the loader, then use format_for_dpo() to prepare data for training.

---

Version: 0.1.0 | Last Updated: December 2025