Dataset · mit-han-lab
The Pile validation split backup
mit-han-lab/pile-val-backup
A backup of the Pile text corpus validation split for language model evaluation.
What's inside
Contains texts from the validation split of the Pile corpus (originally file val.jsonl.zst from the-eye.eu). The dataset size is between 100,000 and 1 million entries. The card notes the requirement to follow the original dataset license.
What you can do with it
- Calculating perplexity and evaluating language model quality
- Calibration during LLM quantization and pruning
- Size
- 100K–1M rows
- Downloads this month
- 19.9K
- Likes
- 27
Models trained on it
- lyf/Huihui-gemma-4-31B-it-abliterated-v2-NVFP4Image text → text 345
- mratsim/GLM-4-32B-0414.w4a16-gptqText generation 117
- Intics/gemma-4-26B-A4B-IT-Int8 90
- EliovpAI/clef-MXFP4Image text → text 66
- BioMistral/BioMistral-7B-AWQ-QGS128-W4-GEMMText generation 63
- BioMistral/BioMistral-7B-AWQ-QGS128-W4-GEMVText generation 28
- mratsim/Seed-OSS-36B-Instruct-NVFP4Text generation 25
- adlik/Qwen3-32B-AWQ-PileText generation 17
How to load it
from datasets import load_dataset
ds = load_dataset("mit-han-lab/pile-val-backup")