Datasets:
license: mit
dataset_info:
features:
- name: clean
dtype: string
- name: corrupted
dtype: string
- name: year
dtype: string
splits:
- name: train
num_bytes: 1234280
num_examples: 10000
download_size: 204638
dataset_size: 1234280
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
language:
- en
Dataset Card for Dataset Name
This is a dataset with examples from the Greater-Than circuit task.
Dataset Details
Dataset Description
- Curated by: Michael Hanna
- Language(s) (NLP): English
- License: MIT
Dataset Sources
- Repository: https://github.com/hannamw/gpt2-greater-than
- Paper: How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model
Uses
This dataset is intended to be a model-agnostic version of the greater-than task.
The original task consisted of examples like The war lasted from the year 1742 to the year 17, based on the fact that GPT-2 small tokenizes 4-digit years into two, two-digit tokens.
One would then compute model performance as the probability assigned to years greater than 42, minus that assigned to years less-than or equal to 42.
New models now tokenize years differently; Llama tokenizes 1742 as [174][2], and Gemma 2 tokenizes it as [1][7][4][2].
You can still compute the probability assigned to good and bad decades; for example:
- For Llama 3, if the token at position [174] is y1, and the token at [2] is y1, you want to compute p(y1>174) + p(y1=174)* p(y2>2) - (p(y1<174) + p(y1=174)* p(y2<=2))
- For Gemma 2, if the token at position [4] is y1, and the token at [2] is y1, you want to compute p(y1>4) + p(y1=4)* p(y2>2) - (p(y1<4) + p(y1=4)* p(y2<=2))
For these purposes, it's easier to have the full string, i.e. The war lasted from the year 1742 to the year 1743, rather than the shortened version The war lasted from the year 1742 to the year 17.
Dataset Structure
clean: The original greater-than example sentences
corrupted: The corrupted version of the corresponding sentence in clean, with the start-year decade set to 01.
year: The start year from the corresponding sentence in clean.
Dataset Creation
Source Data
As described in the paper, this dataset was automatically created, using the template The [event] lasted from the year [XX][YY] to the year [XX].
Michael Hanna and Ollie Liu developed the list of nouns used as [event].
Citation [optional]
BibTeX:
@inproceedings{
hanna2023how,
title={How does {GPT}-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model},
author={Michael Hanna and Ollie Liu and Alexandre Variengien},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023},
url={https://openreview.net/forum?id=p4PckNQR8k}
}
Dataset Card Authors
Michael Hanna