Datasets:
dataset_info:
features:
- name: text
dtype: string
- name: id
dtype: string
- name: metadata
struct:
- name: file_path
dtype: string
- name: input_ids
list: int32
- name: attention_mask
list: int8
splits:
- name: train
num_bytes: 239231368
num_examples: 45736
download_size: 125597135
dataset_size: 239231368
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: odc-by
task_categories:
- text-generation
language:
- en
tags:
- language-modeling
- causal-lm
- llm
size_categories:
- 10K<n<100K
This dataset is a sample of Dolma v1.7 via the 3B version dolma-v1_7-3B. Our sample contains slightly more than 20M tokens (45,736 example texts).
As a pure sample, it maintains the ODC-BY license.
Dataset Description
The columns "id", and "metadata" are copied from the larger dataset, in order to facilitate tracing the source of a particular example.
The columns "input_ids" and "attention_mask" were created with the OLMo tokenizer (a modified version of the GPT-NeoX-20B tokenizer, with some added special tokens). The first token is always "<|endoftext|>".
The original "text" strings are also kept, so users can use another tokenizer if they prefer.
Every example is truncated to at most 1024 tokens (the end is cut off). This affects the "input_ids" (and "attention_mask") column, but not the "text" column. 6791 examples are affected by this.
Curation Rationale
This dataset was primarily created for our project GLUScope, which visualizes strong neuron activations on precisely this dataset. We wanted the dataset to be as lightweight as possible while still providing meaningful information on neuron activations.
Uses
The primary intended use is model analysis work like ours. It is likely to work especially well for OLMo models, since they were trained on Dolma.
However, as with any text dataset, there are many possible use cases. For example, users could use it to train very small language models, do controlled experiments with continued pretraining, and more.
Citation
BibTeX:
[More Information Needed]
Contact
[More Information Needed]