These are GGUF quants of a proof of concept a merge capable of functional 32K context length while being derived from
kukulemon-7B.
The functioning 32K context window has been folded in via a merger of Mistral 7B v0.2 models.
SLERP merge appears to be viable, but DARE-TIES merge risks producing a damaged model and is therefore not recommended.
Although the resulting model natively supports Alpaca prompt, I've tested with ChatML prompts successfuly. Medium temperature (around 1) with low minP (e.g., 0.01) works with ChatML prompts in my most recent testing.
This is a merge of pre-trained language models created using
mergekit.
This model was merged using the SLERP merge method.
1slices:
2 - sources:
3 - model: grimjim/kukulemon-7B
4 layer_range: [0, 32]
5 - model: grimjim/Mistral-7B-Instruct-demi-merge-v0.2-7B
6 layer_range: [0, 32]
7# or, the equivalent models: syntax:
8# models:
9merge_method: slerp
10base_model: grimjim/kukulemon-7B
11parameters:
12 t:
13 - filter: self_attn
14 value: [0, 0.5, 0.3, 0.7, 1]
15 - filter: mlp
16 value: [1, 0.5, 0.7, 0.3, 0]
17 - value: 0.5 # fallback for rest of tensors
18dtype: bfloat16
19
