Dataset · NeelNanda
First 10K texts from The Pile corpus for debugging
NeelNanda/pile-10k
A compact sample of the first 10,000 documents from the well-known The Pile text dataset for rapid testing.
What's inside
A small text dataset containing the first 10,000 items from the famous open corpus The Pile. Prepared by NeelNanda specifically for quick debugging and verification of models trained on the full The Pile, without needing to download the entire data volume.
What you can do with it
- Rapid debugging of training pipelines and evaluation of language models
- Checking perplexity and model behavior on baseline text distributions from The Pile
- Size
- 10K–100K rows
- Downloads this month
- 45.7K
- Likes
- 37
Models trained on it
- derikn/Qwen3.8-27B-AWQ-INT4Image text → text 1.5K
- OPEA/Mistral-Small-3.1-24B-Instruct-2503-int4-AutoRound-awq-sym 1.4K
- OPEA/Qwen2.5-0.5B-Instruct-int4-sym-inc 1.3K
- derikn/Granite-4.2-30B-AWQ-INT4Text generation 667
- bartowski/phi3-4x4b-v1-GGUFText generation 485
- derikn/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-AWQ-INT4Image text → text 386
- mradermacher/phi3-4x4b-v1-GGUF 368
- Intel/gemma-4-31B-it-gguf-q4-k-m-AutoRoundImage text → text 309
How to load it
from datasets import load_dataset
ds = load_dataset("NeelNanda/pile-10k")