Skip to content

Dataset · MJPansa

Native EAGLE features pool for Qwen3.8 Flash-Next

MJPansa/qwen38-eagle4-native-features-37m-20261024-5277c167

Research corpus of 37.7M tokens for training EAGLE-style speculative decoders.

What's inside

Dataset contains 21,082 English dialogues (37,696,946 tokens) and a separate fixed test split. The source corpus is expanded with UltraChat continuations using local FP4-Local/MTP4 while maintaining byte-level accuracy. These are native features for training a Flash-format drafter model, not pre-trained weights or full-vocabulary logit cache.

What you can do with it

  • Training lightweight drafter models for Qwen3.8 speculative decoding
  • Exploring methods to accelerate language model inference
Languages
English
Size
10K–100K rows
Downloads this month
163.9K
Likes
0

How to load it

from datasets import load_dataset

ds = load_dataset("MJPansa/qwen38-eagle4-native-features-37m-20261024-5277c167")