Multi-SpecBench

Multi-SpecBench is a multilingual extension of Spec-Bench, spanning 7 languages × 7 task types, for evaluating speculative decoding and other LLM inference-acceleration methods beyond English. It was introduced in AdaSpec: Adaptive Multilingual Speculative Decoding with Self-Synthesized Language-Aware Training and Vocabulary Simplification (Do, Le, and Nguyen; AAAI-26), a framework that combines language-aware drafter training with adaptive vocabulary simplification for speculative decoding.

Most existing speculative decoding benchmarks (including the original Spec-Bench) are English-only, which hides how well draft models and acceptance rates generalize to other languages and scripts. Multi-SpecBench extends the same task taxonomy to six additional languages and adds a code-mixed multilingual split, so inference-acceleration methods can be measured on non-English and cross-lingual workloads.

Dataset Summary

• 7 languages: English (en), German (de), Spanish (es), French (fr), Japanese (ja), Vietnamese (vi), Chinese (zh) — each with an identical 560-prompt set — plus one additional mix split (588 prompts) sampling across all 7 languages for cross-lingual evaluation.
• 7 task types per language, following the original Spec-Bench taxonomy
• 4,508 prompts total across all 8 files.

Supported Uses

Multi-SpecBench is intended for benchmarking inference-time acceleration methods (vanilla autoregressive decoding, speculative decoding, EAGLE, FR-Spec, AdaSpec, etc.) on a target LLM across languages. Typical usage feeds each prompt to a decoding method under test and measures throughput / acceptance rate / speedup relative to autoregressive decoding, optionally cross-checking output quality against the provided reference fields where available. It is an evaluation-only benchmark — it is not intended for training.

Languages

Config	Language
en	English
de	German
es	Spanish
fr	French
ja	Japanese
vi	Vietnamese
zh	Chinese
mix	Code-mixed sample drawn from all 7 languages above

Dataset Structure

Loading

Each language (plus mix) is exposed as a separate config, each with a single test split:

from datasets import load_dataset

ds = load_dataset("nguyenlab/Multi-SpecBench", "en", split="test")
print(ds[0])

Available configs: en, de, es, fr, ja, vi, zh, mix.

Data Fields

Field	Type	Description
question_id	int	Index of the example in its original source dataset. IDs restart within each category and are not unique across the whole file — treat (category, question_id) as the unique key.
category	string	One of summarization, qa, rag, translation, math_reasoning, code_generation, or one of the 8 MT-Bench-style categories: writing, roleplay, reasoning, math, coding, extraction, stem, humanities.
turns	list[string]	The prompt. Length 1 for the 6 single-turn tasks, length 2 for the 8 MT-Bench-style categories (the second turn is a follow-up instruction).
reference	list[string], optional	Reference answer(s), where available. Present for summarization, rag, translation, math_reasoning, and the MT-Bench categories math, reasoning, coding, extraction. Not present for qa, code_generation, writing, roleplay, stem, humanities. When present alongside a 2-turn prompt, reference has one entry per turn (an empty string if a turn has no judged reference).
lang	string	ISO 639-1 language code, matching the config/file.

Licensing and Attribution

This dataset is released under CC BY-NC 4.0. It is derived from multiple upstream sources (CNN/DailyMail, Natural Questions, DPR, WMT14, GSM8K, HumanEval, MT-Bench) that carry their own licenses and terms of use; please review those upstream licenses before commercial use.

Citation

If you use Multi-SpecBench, please cite both the AdaSpec paper that introduces it and the original Spec-Bench benchmark it builds on:

@article{do2026adaspec,
  title     = {AdaSpec: Adaptive Multilingual Speculative Decoding with Self-Synthesized Language-Aware Training and Vocabulary Simplification},
  author    = {Do, Dinh-Truong and Le, Nguyen-Khang and Nguyen, Le-Minh},
  journal   = {Proceedings of the AAAI Conference on Artificial Intelligence},
  volume    = {40},
  number    = {36},
  pages     = {30530--30538},
  year      = {2026},
  doi       = {10.1609/aaai.v40i36.40307}
}

@inproceedings{xia-etal-2024-unlocking,
  title     = {Unlocking Efficiency in Large Language Model Inference: A Comprehensive Survey of Speculative Decoding},
  author    = {Xia, Heming and Yang, Zhe and Dong, Qingxiu and Wang, Peiyi and Li, Yongqi and Ge, Tao and Liu, Tianyu and Li, Wenjie and Sui, Zhifang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2024},
  month     = aug,
  year      = {2024},
  pages     = {7655--7671}
}

Acknowledgments

Multi-SpecBench builds directly on Spec-Bench and its underlying task datasets (CNN/DailyMail, Natural Questions, DPR, WMT14, GSM8K, HumanEval, MT-Bench).