D3D
Collection
Decoupled Diffusion Drafting for Asynchronous Speculative Decoding • 15 items • Updated • 3
Datasets:
Star the D3D repository to request access
Please star https://github.com/fudan-generative-vision/D3D before requesting access.
I have starred the D3D repository.
Log in or Sign Up to review the conditions and access this dataset content.
PerfectBlend: Qwen3-4B NoThink (Sampling)
Overview
This dataset contains 1,420,909 prompt–response records generated by Qwen3-4B on the Open-PerfectBlend prompts. Responses use non-thinking mode and sampling decoding.
Generation recipe
| Item | Value |
|---|---|
| Generator | Qwen/Qwen3-4B |
| Prompt source | mlabonne/open-perfectblend (1,420,909 prompts) |
| Mode | Non-thinking |
| Decoding | Sampling |
| Context length | 40,960 |
max_tokens |
4,096 |
| Records | 1,420,909 |
Sampling parameters
| Parameter | Value |
|---|---|
temperature |
0.7 |
top_p |
0.8 |
top_k |
20 |
min_p |
0.0 |
presence_penalty |
0.0 |
repetition_penalty |
1.0 |
Data format
| Field | Description |
|---|---|
id |
Source prompt row index in Open-PerfectBlend. |
conversations |
Messages with role and content. |
status |
Generation status. |
Loading
hf auth login
hf download fudan-generative-ai/PerfectBlend-Qwen3-4B-NoThink-Sampling \
perfectblend_qwen3-4b-nothink-sampling_regen.jsonl \
--repo-type dataset --local-dir ./perfectblend-qwen3-4b-nothink-sampling
import json
path = "./perfectblend-qwen3-4b-nothink-sampling/perfectblend_qwen3-4b-nothink-sampling_regen.jsonl"
with open(path, encoding="utf-8") as f:
for line in f:
record = json.loads(line)
messages = record["conversations"]
License
Dataset: Apache-2.0.
Citation
@article{chen2026d3d,
title = {D3D: Decoupled Diffusion Drafting for Asynchronous Speculative Decoding},
author = {Baoyou Chen and Ruiqiao Mei and Guancong Lin and Tao Song and Weihao Yuan and Qun Liu and Siyu Zhu},
journal = {arXiv preprint arXiv:2611.06425},
year = {2026}
}
@article{xu2024perfectblend,
title = {The Perfect Blend: Redefining RLHF with Mixture of Judges},
author = {Xu, Tengyu and Helenowski, Eryk and Sankararaman, Karthik Abinav and Jin, Di and Peng, Kaiyan and Han, Eric and Nie, Shaoliang and Zhu, Chen and Zhang, Hejia and Zhou, Wenxuan and Zeng, Zhouhao and He, Yun and Mandyam, Karishma and Talabzadeh, Arya and Khabsa, Madian and Cohen, Gabriel and Tian, Yuandong and Ma, Hao and Wang, Sinong and Fang, Han},
journal = {arXiv preprint arXiv:2409.20370},
year = {2024}
}
Sources:
- D3D paper: https://arxiv.org/abs/2611.06425
- D3D code and data: https://github.com/fudan-generative-vision/D3D
- Qwen3-4B: https://huggingface.co/Qwen/Qwen3-4B
- Open-PerfectBlend: https://huggingface.co/datasets/mlabonne/open-perfectblend
- The Perfect Blend: https://arxiv.org/abs/2409.20370
- Downloads last month
- 6