data

data/

Bundled calibration and evaluation datasets. Committed to the repository. No internet needed after scripts/prepare_data.py is run once.

Structure

data/
├── calibration/          # 128 samples per dataset, for AWQ/GPTQ calibration
│   ├── wikitext2_128.jsonl
│   ├── alpaca_128.jsonl
│   ├── gsm8k_128.jsonl
│   ├── humaneval_128.jsonl
│   ├── qa_128.jsonl
│   ├── sharegpt_128.jsonl
│   └── sum_128.jsonl
└── eval/                 # Full eval splits, for PPL measurement
    ├── wikitext2_eval.jsonl
    ├── alpaca_eval.jsonl
    ├── gsm8k_eval.jsonl
    ├── humaneval_eval.jsonl
    ├── qa_eval.jsonl
    ├── sharegpt_eval.jsonl
    └── sum_eval.jsonl

File Format

Each .jsonl file has one JSON object per line: {"text": "..."}.

Dataset Sources

Name	HuggingFace Source	Domain
wikitext2	Salesforce/wikitext, wikitext-2-raw-v1	Wikipedia prose
alpaca	tatsu-lab/alpaca	Instruction following
gsm8k	openai/gsm8k, main split	Math word problems
humaneval	openai/openai_humaneval	Python programming
qa	rajpurkar/squad	Reading comprehension
sharegpt	lmsys/chatbot_arena_conversations	Multi-turn chat
sum	EdinburghNLP/xsum	News summarization

Why Bundle Datasets?

1. Reproducibility — same calibration data every time, regardless of HF dataset updates
2. Offline eval — run nanoptq eval without internet after initial setup
3. Student-friendly — clone and run, no accounts or API keys needed

Regenerating

python scripts/prepare_data.py                           # all datasets
python scripts/prepare_data.py --dataset gsm8k alpaca   # specific datasets