Quantization made by Richard Erkhov.
This is a merge of pre-trained language models created using
mergekit.
We explore merger at extremely low weight as an alternative to fine-tuning. The additional model was applied at a weight of 10e-5, which was selected to be comparable to a few epochs of training. The low weight also amounts to the additional model being flattened, though technically not sparsified.
This model was merged using the
task arithmetic merge method using
grimjim/kukulemon-32K-7B as a base.
1base_model: grimjim/kukulemon-32K-7B
2dtype: bfloat16
3merge_method: task_arithmetic
4slices:
5- sources:
6 - layer_range: [0, 32]
7 model: grimjim/kukulemon-32K-7B
8 - layer_range: [0, 32]
9 model: grimjim/rogue-enchantress-32k-7B
10 parameters:
11 weight: 10e-5
12
