fffoivos's picture
docs: describe preference datasets without assigning a training method
fc13ac4 verified
|
Raw History Blame Contribute Delete
3.74 kB
metadata
language:
  - el
  - en
  - fr
  - de
  - es
  - it
  - pt
pretty_name: Greek Apertus 8B, post-training data
license: other
license_name: mixed-per-part
license_link: >-
  https://huggingface.co/datasets/glossAPI/greek-apertus-post-training-data/blob/main/README.md
extra_gated_prompt: >-
  The post-training data of Greek Apertus 8B: second-stage SFT, chat SFT and
  preference pairs. Contains third-party data under its own licences alongside
  the project's own Greek sets. By requesting access you agree to use the data
  for research and to respect the licences of each part.
extra_gated_fields:
  Affiliation: text
  Intended use: text
tags:
  - apertus
  - greek
  - glossapi
  - sft
  - preference-data
  - chat

Greek Apertus 8B: post-training data

Second-stage SFT, chat SFT and preference pairs.

Contents

Stage Data
Greek maths 5,482 maths rows + 4,496 replay rows
Conversation 3,185 chat turns + 568 replay rows
Preference datasets dpo/

Repositories and checkpoints

Licences

Mixed, per part; each part's original card states its own terms.

Acknowledgements

This work was implemented thanks to a grant by Swiss AI for compute on CSCS.

Collection: Greek Apertus 8B.