How to obtain the calibration dataset used in your experiments?

[BrotherHappy]

Hi, thanks for your great work and for open-sourcing this project!

I have a question regarding the calibration dataset used in your experiments. For reproducibility, could you please clarify:

• What exact dataset (or data source) was used for calibration?
• How was the dataset constructed (e.g., sampling strategy, preprocessing, tokenization, sequence length, etc.)?
• Is there an official way to download or reproduce the same calibration dataset?

It would be extremely helpful if you could provide a download link or detailed instructions to reconstruct the dataset.

Thanks in advance for your help!

[localai-org-maint-bot]

This is a valid reproducibility gap. The exact calibration source, sampling/preprocessing, tokenization, and sequence length should be versioned or scripted alongside the recipes; without that, published quality claims cannot be independently reproduced. @mudler this should be treated as documentation/data provenance work, and it also blocks clean support answers such as #16.

[nmorgowicz]

Can i bump this too please? i want to be able to APEX quant models as well, and while i can use pre-existing imatrix files from mradermacher, it's not the same as having the dataset that is used as part of the APEX quantization you use on models posted up on HF.

@mudler can you please help or guide those of us who want custom private quantizations of finetunes to also be APEX'ified to be able to do so with the proper data?