|
--- |
|
configs: |
|
- config_name: 0k |
|
data_files: |
|
- split: qa1 |
|
path: data/qa1/0k.json |
|
- split: qa2 |
|
path: data/qa2/0k.json |
|
- split: qa3 |
|
path: data/qa3/0k.json |
|
- split: qa4 |
|
path: data/qa4/0k.json |
|
- split: qa5 |
|
path: data/qa5/0k.json |
|
- split: qa6 |
|
path: data/qa6/0k.json |
|
- split: qa7 |
|
path: data/qa7/0k.json |
|
- split: qa8 |
|
path: data/qa8/0k.json |
|
- split: qa9 |
|
path: data/qa9/0k.json |
|
- split: qa10 |
|
path: data/qa10/0k.json |
|
- config_name: 1k |
|
data_files: |
|
- split: qa1 |
|
path: data/qa1/1k.json |
|
- split: qa2 |
|
path: data/qa2/1k.json |
|
- split: qa3 |
|
path: data/qa3/1k.json |
|
- split: qa4 |
|
path: data/qa4/1k.json |
|
- split: qa5 |
|
path: data/qa5/1k.json |
|
- split: qa6 |
|
path: data/qa6/1k.json |
|
- split: qa7 |
|
path: data/qa7/1k.json |
|
- split: qa8 |
|
path: data/qa8/1k.json |
|
- split: qa9 |
|
path: data/qa9/1k.json |
|
- split: qa10 |
|
path: data/qa10/1k.json |
|
- config_name: 2k |
|
data_files: |
|
- split: qa1 |
|
path: data/qa1/2k.json |
|
- split: qa2 |
|
path: data/qa2/2k.json |
|
- split: qa3 |
|
path: data/qa3/2k.json |
|
- split: qa4 |
|
path: data/qa4/2k.json |
|
- split: qa5 |
|
path: data/qa5/2k.json |
|
- split: qa6 |
|
path: data/qa6/2k.json |
|
- split: qa7 |
|
path: data/qa7/2k.json |
|
- split: qa8 |
|
path: data/qa8/2k.json |
|
- split: qa9 |
|
path: data/qa9/2k.json |
|
- split: qa10 |
|
path: data/qa10/2k.json |
|
- config_name: 4k |
|
data_files: |
|
- split: qa1 |
|
path: data/qa1/4k.json |
|
- split: qa2 |
|
path: data/qa2/4k.json |
|
- split: qa3 |
|
path: data/qa3/4k.json |
|
- split: qa4 |
|
path: data/qa4/4k.json |
|
- split: qa5 |
|
path: data/qa5/4k.json |
|
- split: qa6 |
|
path: data/qa6/4k.json |
|
- split: qa7 |
|
path: data/qa7/4k.json |
|
- split: qa8 |
|
path: data/qa8/4k.json |
|
- split: qa9 |
|
path: data/qa9/4k.json |
|
- split: qa10 |
|
path: data/qa10/4k.json |
|
- config_name: 8k |
|
data_files: |
|
- split: qa1 |
|
path: data/qa1/8k.json |
|
- split: qa2 |
|
path: data/qa2/8k.json |
|
- split: qa3 |
|
path: data/qa3/8k.json |
|
- split: qa4 |
|
path: data/qa4/8k.json |
|
- split: qa5 |
|
path: data/qa5/8k.json |
|
- split: qa6 |
|
path: data/qa6/8k.json |
|
- split: qa7 |
|
path: data/qa7/8k.json |
|
- split: qa8 |
|
path: data/qa8/8k.json |
|
- split: qa9 |
|
path: data/qa9/8k.json |
|
- split: qa10 |
|
path: data/qa10/8k.json |
|
- config_name: 16k |
|
data_files: |
|
- split: qa1 |
|
path: data/qa1/16k.json |
|
- split: qa2 |
|
path: data/qa2/16k.json |
|
- split: qa3 |
|
path: data/qa3/16k.json |
|
- split: qa4 |
|
path: data/qa4/16k.json |
|
- split: qa5 |
|
path: data/qa5/16k.json |
|
- split: qa6 |
|
path: data/qa6/16k.json |
|
- split: qa7 |
|
path: data/qa7/16k.json |
|
- split: qa8 |
|
path: data/qa8/16k.json |
|
- split: qa9 |
|
path: data/qa9/16k.json |
|
- split: qa10 |
|
path: data/qa10/16k.json |
|
- config_name: 32k |
|
data_files: |
|
- split: qa1 |
|
path: data/qa1/32k.json |
|
- split: qa2 |
|
path: data/qa2/32k.json |
|
- split: qa3 |
|
path: data/qa3/32k.json |
|
- split: qa4 |
|
path: data/qa4/32k.json |
|
- split: qa5 |
|
path: data/qa5/32k.json |
|
- split: qa6 |
|
path: data/qa6/32k.json |
|
- split: qa7 |
|
path: data/qa7/32k.json |
|
- split: qa8 |
|
path: data/qa8/32k.json |
|
- split: qa9 |
|
path: data/qa9/32k.json |
|
- split: qa10 |
|
path: data/qa10/32k.json |
|
--- |
|
|
|
# BABILong (5k train samples) : a long-context needle-in-a-haystack benchmark for LLMs |
|
|
|
Preprint is on [arXiv](https://arxiv.org/abs/2402.10790) |
|
|
|
## bAbI + Books = BABILong |
|
|
|
**BABILong** is a novel generative benchmark for evaluating the performance of NLP models in |
|
processing arbitrarily long documents with distributed facts. |
|
|
|
It contains 10 configs, each corresponding to its bAbI task. Each config has spltis corresponding to different sequence lengths in tokens: '4k', '32k', '128k', '256k', '512k', '1M' |
|
|
|
Solving tasks with a long context size requires the model to distinguish important information from large amounts of irrelevant details. To simulate this behavior we ”hide” the sentences of the original task between the sentences of irrelevant text. We use the [bAbI](https://huggingface.co/datasets/facebook/babi_qa) dataset [1] as facts and [PG19](https://huggingface.co/datasets/pg19) as background text. Resulting test samples might have lenghts of **millions of tokens**. |
|
|
|
BABILong consists of 10 tasks designed for evaluation of basic aspects of reasoning. The bAbI tasks are generated by simulating a set of characters and objects engaged in various movements and interactions with each other in multiple locations. Each interaction is represented by a fact, e.g. **”Mary travelled to the office”**, and the task is to answer a question using the facts from the current simulation, for instance, **”Where is Mary?”**. The bAbI tasks vary based on the number of facts, question complexity and the aspects of reasoning. |
|
|
|
### First ten tasks of BABILong |
|
|
|
| Task | Name | facts per task | supporting facts per task | |
|
|------|--------------------------|-----------------|---------------------------| |
|
| qa1 | single supporting fact | 2 - 10 | 1 | |
|
| qa2 | two supporting facts | 2 - 68 | 2 | |
|
| qa3 | three supporting facts | 4 - 32 | 3 | |
|
| qa4 | two arg relations | 2 | 1 | |
|
| qa5 | three arg relations | 2 - 126 | 1 | |
|
| qa6 | yes-no questions | 2 - 26 | 1 | |
|
| qa7 | counting | 2 - 52 | 1-10 | |
|
| qa8 | lists-sets | 2 - 50 | 1-8 | |
|
| qa9 | simple negation | 2 - 10 | 1 | |
|
| qa10 | indefinite knowledge | 2 - 10 | 1 | |
|
|
|
Join us in this exciting endeavor and let's push the boundaries of what's possible together! |
|
|
|
## Citation |
|
``` |
|
@misc{kuratov2024search, |
|
title={In Search of Needles in a 10M Haystack: Recurrent Memory Finds What LLMs Miss}, |
|
author={Yuri Kuratov and Aydar Bulatov and Petr Anokhin and Dmitry Sorokin and Artyom Sorokin and Mikhail Burtsev}, |
|
year={2024}, |
|
eprint={2402.10790}, |
|
archivePrefix={arXiv}, |
|
primaryClass={cs.CL} |
|
} |
|
``` |
|
|
|
## References |
|
|
|
[1] Weston, Jason, et al. "Towards ai-complete question answering: A set of prerequisite toy tasks." arXiv preprint [arXiv:1502.05698](https://arxiv.org/abs/1502.05698) (2015). |
