synthseq/flipflop

Data for Flip-Flop Language Modeling. The task is to correctly execute the sequential operations of a 1-bit register. The Transformer architecture, despite being apparently built for this operation, makes sporadic extrapolation errors (attention glitches). An open challenge is to fix these without recourse to long-tailed data or a recurrent architecture. Splits reflect the FFLM setup from the paper:

• train: 1.6M sequences from FFL(0.8) (256 instructions, 80% ignore, 10% read, 10% write).
• val: 16K sequences from FFL(0.8).
• val_dense: 4K sequences from FFL(0.1).
• val_sparse: 160K sequences from FFL(0.98).

Usage

import torch
import datasets

dataset = datasets.load_dataset('synthseq/flipflop')
dataset['train'][0]  # {'text': 'w1i1w0i0 ...

def tokenize_batch(batch):
    mapping = {'w': 0, 'r': 1, 'i': 2, '0': 3, '1': 4}
    tokenized_batch = [[mapping[char] for char in s] for s in batch['text']]
    return {'tokens': torch.tensor(tokenized_batch, dtype=torch.int64)}

dataset.set_transform(tokenize_batch)
dataset['train'][0]  # {'tokens': tensor([0, 4, 2, 4, 0, 3, 2, 3, 2 ...

Citation

@article{liu2023exposing,
  title={Exposing Attention Glitches with Flip-Flop Language Modeling},
  author={Liu, Bingbin and Ash, Jordan T and Goel, Surbhi and Krishnamurthy, Akshay and Zhang, Cyril},
  journal={arXiv preprint arXiv:2306.00946},
  year={2023}
}