metadata
license: apache-2.0
task_categories:
- text-generation
- question-answering
language:
- en
tags:
- DCLM
- fineweb
- edu
- pretrain
- pretraining
- web
- reason
- reasoning
- dense
- fine
- fine data
- ultra
- pro
- DCLM Pro
- DCLM baseline
- thought
- thoughts
- think
- thinking
- post train
- posttrain
size_categories:
- 1M<n<10M
DCLM Ultra
DCLM Ultra is an improved, deduplicated and filtered version of DCLM Pro (Filtered DCLM, 1.5% better compared to DCLM baseline).
It achieves superior performance across benchmarks including hellaswag and piqa, also achieving better loss compared to pro.
3M model trained on only DCLM Pro vs only Ultra, same architecture, same training budget.
| Metric | DCLM Pro | DCLM Ultra |
|---|---|---|
| Training loss | 3.37 | 3.27 |
| PIQA | 53.65% | 54.95% |
| HellaSwag | 25.45% | 25.54% |
Dataset Details
- Curated by: AInion
- Funded by: ArrowAI
- Shared by: MoreThought
- Language: english
- License: apache 2.0
Dataset Sources
- Main Source: https://huggingface.co/datasets/mlfoundations/dclm-baseline-1.0
- Pro Source: https://huggingface.co/datasets/mlfoundations/dclm-baseline-1.0-parquet
- Pro Repo: https://huggingface.co/datasets/gair-prox/DCLM-pro
Uses
- PreTraining SLMs on clean, heavily filtered data.
- PostTraining LLMs on clean, heavily filtered data.
- PreTraining data quality benefits on SLMs research.