RULER, DeepSeek-V3.2 tokenizer
Synthetic long-context tasks generated with NVIDIA RULER. Token lengths are measured with the DeepSeek-V3.2 tokenizer. Seed 20260823, 20 examples per task.
Six lengths: 4k, 8k, 16k, 32k, 64k, 128k. Thirteen tasks at each length, 260 examples per length, 1560 total.
Tasks: niah_single_1, niah_single_2, niah_single_3, niah_multikey_1, niah_multikey_2, niah_multikey_3, niah_multiquery, niah_multivalue, vt, cwe, fwe, qa_1, qa_2.
Each validation.jsonl row has index, input, outputs, length, length_w_model_temp, answer_prefix, and token_position_answer. Generation settings are in generation_metadata.json.
@article{hsieh2024ruler,
title={RULER: What's the Real Context Size of Your Long-Context Language Models?},
author={Cheng-Ping Hsieh and Simeng Sun and Samuel Kriman and Shantanu Acharya and Dima Rekesh and Fei Jia and Yang Zhang and Boris Ginsburg},
journal={arXiv preprint arXiv:2404.06654},
year={2024}
}
- Downloads last month
- 42
Paper for Yanlin77/ruler
Paper • 2404.06654 • Published • 43