The model did achieve an improvement in TruthfulQA over cookinai/CatMacaroni-Slerp and GSM8K over mncai/mistral-7b-dpo-v5 which was the goal of the merge leading to an average score that was a better than both. It is unclear why the TruthfulQA metric is still meaningfully lower than the base mncai/mistral-7b-dpo-v5.

Training Details

.yaml file for mergekit

yaml

slices:
  - sources:
      - model: cookinai/CatMacaroni-Slerp
        layer_range: [0, 32]
      - model: mncai/mistral-7b-dpo-v5
        layer_range: [0, 32]
merge_method: slerp
base_model: mncai/mistral-7b-dpo-v5
parameters:
  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
dtype: float16

Bias, Risks, and Limitations

The model has not been evaluated for safety and is only intended for research and experiments.

abacusai/Slerp-CM-mist-dpo

Evaluation Results

HuggingFace Leaderboard

Training Details

Bias, Risks, and Limitations

abacusai/
Slerp-CM-mist-dpo