models: | |
- model: allenai/tulu-2-dpo-7b | |
parameters: | |
weight: 1.0 | |
# - model: EleutherAI/llemma_7b | |
# parameters: | |
# weight: 1.0 | |
merge_method: svd_franken_merge | |
base_model: meta-llama/Llama-2-7b-chat-hf | |
parameters: | |
probabilistic: True # use probabilistic SVD algorithm (maybe fastest but little inaccurate) instead of the non probabilistic SVD algorithm (slowest but accurate) | |
sv_reduction: 1.0 # strict value: "1.0 / number of task vector" if number of task vector is >=2 (percentage of singular components to keep for each TV) | |
sv_scaling: 1.0 # float (hyperparameter): suggested nearby "number of task vector / 2.0" if number of task vector is >=2 else 1.0 (reduction to the stretching factor of the singular values) | |
num_iterations: 4 # number of iterations for the probabilistic SVD algorithm (2 is a good start and fastest option but could be inaccurate, 32 is the most accurate but slowest option. I don't recommend going above 32 use SVD probabilistic at False instead) | |
dtype: float16 | |