metadata
base_model:
- swiss-ai/Apertus-8B-2509
tags:
- text-generation-inference
- transformers
- unsloth
- apertus
- trl
- cpo
license: apache-2.0
language:
- en
Actually functional this time Alpha/SimPO checkpoint trained from Apertus base.
Trained on a mix of curated C2, Gutenberg, and Instruct Skill-Mix
Alpaca chat template, temp .5, min_p .05, rep pen 1.05 seems reasonable.
Now I just have to make a better preference dataset...