TheProfessor-155b / mergekit_config.yml
ehartford's picture
Upload folder using huggingface_hub
a11426d verified
raw
history blame
1.97 kB
merge_method: linear # use linear so we can include multiple models, albeit at a zero weight
parameters:
weight: 1.0 # weight everything as 1 unless specified otherwise - linear with one model weighted at 1 is a no-op like passthrough
slices:
- sources:
- model: cognitivecomputations/dolphin-2.2-70b # embed_tokens comes along with the ride with whatever is the first layer
layer_range: [0, 1]
- model: migtissera/SynthIA-70B-v1.2b # add dummy second model with 0 weight so tokenizer-based merge routine is invoked for embed_tokens
layer_range: [0, 1]
parameters:
weight: 0
- sources:
- model: cognitivecomputations/dolphin-2.2-70b
layer_range: [1, 20]
- sources:
- model: migtissera/SynthIA-70B-v1.2b
layer_range: [10, 30]
- sources:
- model: WizardLM/WizardMath-70B-V1.0
layer_range: [20, 40]
- sources:
- model: epfl-llm/meditron-70b
layer_range: [25, 45]
- sources:
- model: cognitivecomputations/dolphin-2.2-70b
layer_range: [30, 50]
- sources:
- model: migtissera/SynthIA-70B-v1.2b
layer_range: [40, 60]
- sources:
- model: WizardLM/WizardMath-70B-V1.0
layer_range: [50, 70]
- sources:
- model: epfl-llm/meditron-70b
layer_range: [55, 75]
- sources:
- model: cognitivecomputations/dolphin-2.2-70b
layer_range: [60, 79]
- sources: # same as above, but for lm_head with the last layer
- model: cognitivecomputations/dolphin-2.2-70b
layer_range: [79, 80]
- model: migtissera/SynthIA-70B-v1.2b
layer_range: [79, 80]
parameters:
weight: 0
dtype: float16
tokenizer_source: model:cognitivecomputations/dolphin-2.2-70b # keep exact tokenizer used by dolphin - or you could use `union` if you add all of the input models to the first/last slice, but they would need to be non-zero weight or you'll get NaNs in your embeddings