This merge is intended as an intermediate merge for further merges. It's useable as-is, and doesn't show any glaring signs of broken behavior. I've included a Q8_0 gguf in the repo if anyone is curious to try it.

The intuition behind this merge is as follows:

We keep the top 65 percent of weight deltas from the WPO-HB fine tune at a (very near) 1.0 weight We 'flood fill' the remaining 45 percent of model weights with the SimPO weights. Because normalize is set to true, this results in a 1.0 weight from SimPO's deltas on any non-overlapping weights. In cases of overlap with the top 65 weights from WPO-HB, the relative weight contribution for SimPO is near zero. In cases of non-overlap, each model gets a 1.0 or very near 1.0 weight for the merge.

Per the mergekit docs:

weight - relative (or absolute if normalize=False) weighting of a given tensor
normalize - if true, the weights of all models contributing to a tensor will be normalized. Default behavior.

🤔 Assuming I've understood correctly, anyway. If anyone knows of a way to get model stats, ex, something like a diff that just spits out stats about the weight differences between two models, feel free to leave a comment in the discussion. I'd love to be able to verify exactly what's happened to the weights, and if it's matching my expectations 🤷

This is a merge of pre-trained language models created using mergekit.

This model was merged using the della merge method using google/gemma-2-9b-it as a base.

The following models were included in the merge:

The following YAML configuration was used to produce this model: