Hugging Face's logo

Datasets:
wingchiuloong
/
specmix

Modalities:
Text
Formats:
parquet
Size:
1K - 10K
ArXiv:
Tags:
evaluation
benchmark
mixed-traffic
speculative-decoding
large-language-models
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
specmix / README.md
wingchiuloong's picture
wingchiuloong
Fix references: link to DCore paper and public code repo
2d6038d verified
preview code
|
Raw History Blame Contribute Delete
4.73 kB
metadata 
license: cc-by-nc-sa-4.0
tags:
  - evaluation
  - benchmark
  - mixed-traffic
  - speculative-decoding
  - large-language-models
pretty_name: SpecMix (Mixed-Traffic Speculative Decoding Evaluation Set)
size_categories:
  - 1K<n<10K

SpecMix

A fixed, shuffled mixed-traffic evaluation set for speculative decoding (draft-then-verify) serving. SpecMix interleaves requests from 10 public benchmarks the way a real serving endpoint sees traffic, so throughput at high concurrency can be measured on a realistic mixture instead of one homogeneous domain at a time. It is the mixed-traffic benchmark used in the DCore evaluation pipeline.

Composition

# Source Rows Source license Source (Hugging Face)
1 gsm8k 500 MIT openai/gsm8k
2 math500 500 MIT (via prm800k) HuggingFaceH4/MATH-500
3 aime25 30 CC-BY-NC-SA-4.0 MathArena/aime_2025
4 humaneval 164 MIT openai/openai_humaneval
5 mbpp 257 CC-BY-4.0 google-research-datasets/mbpp
6 livecodebench 500 gated (terms apply) livecodebench/livecodebench
7 mt-bench 80 NC (see below) lmsys/mt-bench
8 alpaca 500 CC-BY-NC-4.0 tatsu-lab/alpaca
9 arena-hard 500 Apache-2.0 (card); real user queries lmarena-ai/arena-hard-auto-v0.1
10 swe-bench 300 MIT princeton-nlp/SWE-bench_Lite
Total 3331

Rows are the first N of each source (per the quota above), merged and shuffled with a fixed seed (42), so the mix is frozen and reproducible.

Fields

  • turns — single-element list with the already-formatted prompt (identical string to what the per-dataset evaluation path produces).
  • source — originating dataset name.
  • source_index — row index inside that dataset (for cross-referencing).
  • gold — gold answer for scorable sources (gsm8k, math500, aime25; 1,030 rows), "" otherwise. Gold answers are numeric only; no code or patch content is included in this dataset.

Usage 

from datasets import load_dataset
ds = load_dataset("wingchiuloong/specmix", split="train")
print(ds[0])

Regeneration

This dataset is generated by scripts/build_specmix_dataset.py in the ernie-research/DCore repository:

python scripts/build_specmix_dataset.py   # writes cache/dataset/specmix (seed=42)

License and terms

  • This combined dataset is distributed under CC-BY-NC-SA-4.0 (the most restrictive license among its sources, determined by aime25 CC-BY-NC-SA-4.0 and alpaca/mt-bench CC-BY-NC-4.0). Research / non-commercial use only.
  • Each component remains under its original license; refer to the source links above for full terms. mbpp (CC-BY-4.0) requires attribution, which is provided here.
  • aime25 problem statements are © the Mathematical Association of America (MAA), as released by MathArena.
  • livecodebench, lmsys/mt-bench and lmsys/lmsys-chat-1m are gated on Hugging Face and subject to their own terms of use. arena-hard rows are real user queries from Chatbot Arena; they may contain personal data. Please respect the terms of the underlying sources when using the corresponding rows.
  • If you are a rights holder of any source and would like this dataset removed or adjusted, please open an issue on the ernie-research/DCore repository.