My first successful Dare-Ties merge. Because of the tokenizer difference of the model types (also bf16 vs f16), Had to use Slerp as well.
Seems to perform well! Did a local lm-eval and HellaSWAG gives me around 84.5, which seems decent. will be submitting this for eval on the openLLM leaderboard as well.
Preset for this should be ChatML, but standard default presets should work ok too.
base_model:
- senseable/WestLake-7B-v2
- cognitivecomputations/dolphin-2.8-mistral-7b-v02 library_name: transformers tags:
- mergekit
- merge
Noodlz_DolphinLake-DARE_TIE_SLERP-tokenwest
This is a merge of pre-trained language models created using mergekit.
Merge Details
Merge Method
This model was merged using the DARE TIES merge method using cognitivecomputations/dolphin-2.8-mistral-7b-v02 as a base.
Models Merged
The following models were included in the merge:
Configuration
The following YAML configuration was used to produce this model:
merge_method: dare_ties
parameters:
int8_mask: true
t:
- filter: self_attn
value: [0, 0.5, 0.3, 0.7, 1]
- filter: mlp
value: [1, 0.5, 0.7, 0.3, 0]
- value: 0.5 # fallback for rest of tensors
embed_slerp: true
models:
- model: cognitivecomputations/dolphin-2.8-mistral-7b-v02
# No parameters necessary for base model
- model: senseable/WestLake-7B-v2
parameters:
density: 0.58
weight: 0.8
base_model: cognitivecomputations/dolphin-2.8-mistral-7b-v02
tokenizer_source: model:senseable/WestLake-7B-v2
dtype: bfloat16