What was the training dataset/data mix used to train AngelSlim/Qwen3-8B_eagle3 and other qwen eagle3 drafters?

[nelaturuharsha]

Hi,

I was looking to run some experiments and was curious what data was used to train AngelSlim/Qwen3-8B_eagle3 - was it just ultrachat200k or sharegpt or a mix?

Is this information documented/available somewhere? Thanks in advance.

Edit: I saw in other issues that both ultrachat200k and shareGPT were used. I am curious what the exact splits were

i.e. was it full concatenation of https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k and https://huggingface.co/datasets/Aeala/ShareGPT_Vicuna_unfiltered (i.e. 200k samples from ultrachat and 121+k from ShareGPT or just 68k for ShareGPT)

More details on this would be really helpful, thanks!

[liusong1222]

Hi,

The training data is a mixture of UltraChat and ShareGPT, using full concatenation of both datasets (i.e., all available samples from each were included).