speculators
qwen3-8b-peagle-speculatorsdflash-qwen3-8b-speculatorsQwen3-8B-from-Qwen3-8B_regen-speculators.eagle3-qwen3arch-ckpt1Qwen3-8B-speculators.peagle-qwen3arch-ckpt4qwen3-8b.speculators.dspark-ce01tv09-bs8-conf-swallama3.1-8b-it.speculators.dflash-ce01tv09-bs8-swallama3.1-8b-it.speculators.dspark-ce01tv09-bs8-conf-swaqwen3-8b.speculators.dflash-ce01tv09-bs8-swa
speculators-ci-datasets
speculator-tutorial
Raw vs. on-policy regenerated conversation data for training speculative-decoding
drafters (EAGLE-3 / DFlash / DSpark style), with the original source data kept alongside
so you can see exactly what regeneration changes and why it matters.
Prompts come from UltraChat-200k. The verifier / teacher model is Qwen/Qwen3-8B.
Why regenerate at all?
A speculative-decoding drafter is trained to predict what the verifier would say next.
If you train it… See the full description on the dataset page: https://huggingface.co/datasets/inference-optimization/speculators-ci-datasets.speculators_benchmarks_tool_callspeculators-multilingual-en-fr-de-it-es
Speculators Multilingual SFT Dataset (en/fr/de/it/es)
A multilingual instruction-following dataset in ShareGPT format, built to train draft models for speculative decoding across English, French, German, Italian and Spanish.
Summary
An English instruction-tuning corpus with part of it kept in English and the rest machine-translated into French, German, Italian and Spanish using tencent/Hunyuan-MT-7B. Provided as a single mixed-language, ShareGPT-formatted dataset… See the full description on the dataset page: https://huggingface.co/datasets/Infomaniak-AI/speculators-multilingual-en-fr-de-it-es.speculators-multilingual-en-fr-de-it-es
Speculators Multilingual SFT Dataset (en/fr/de/it/es)
A multilingual instruction-following dataset in ShareGPT format, built to train draft models for speculative decoding across English, French, German, Italian and Spanish.
Summary
An English instruction-tuning corpus with part of it kept in English and the rest machine-translated into French, German, Italian and Spanish using tencent/Hunyuan-MT-7B. Provided as a single mixed-language, ShareGPT-formatted dataset… See the full description on the dataset page: https://huggingface.co/datasets/ulysse1/speculators-multilingual-en-fr-de-it-es.speculators-qwen3-30b-a3b-instruct-2507speculators-example-data
Example Data Format
This directory contains example data showing the format used in the speculators project.
Directory Structure
example_data/
├── preprocessed_data/ # INPUT: Arrow format (HuggingFace Dataset)
│ ├── data-00000-of-00001.arrow
│ ├── dataset_info.json
│ └── state.json
└── hidden_states/ # OUTPUT: Safetensors format
└── cmpl-*.safetensors
Quick Inspection Commands
View Preprocessed Data (Arrow format)
#… See the full description on the dataset page: https://huggingface.co/datasets/shanjiaz/speculators-example-data.