metadata
dataset_info:
features:
- name: raw_content
dtype: string
- name: input_ids
sequence: int32
- name: attention_mask
sequence: int8
- name: labels
sequence: int64
splits:
- name: train
num_bytes: 6760321689
num_examples: 32000
download_size: 2955669157
dataset_size: 6760321689
Dataset Card for "context_extension-mistral-natural_distribution-16k"
- 32k samples
- mistral token size 1 < x < 16400
- ~natural size distribution (lots of small, few long, 1/x-like)
- build with small (no filter) 1/3 + 2/3 of long (+16k), resize all +6k to (1<x<16_400)
- from redpajama-v2