Konstanta-Alpha-V2-7B

This is a merge of pre-trained language models created using mergekit.

Merge Details

Merge Method

This model was merged using the DARE TIES to merge Kunoichi with PiVoT Evil and to merge ArchBeagle with Silicon Alice, and then merge the resulting 2 models with the gradient SLERP merge method. ChatML seems to work the best.

Models Merged

The following models were included in the merge:

• SanjiWatsuki/Kunoichi-DPO-v2-7B
• maywell/PiVoT-0.1-Evil-a
• mlabonne/ArchBeagle-7B
• LakoMoor/Silicon-Alice-7B

Configuration

The following YAML configuration was used to produce this model (to reproduce use mergekit-mega command):

base_model: mistralai/Mistral-7B-v0.1
dtype: float16
merge_method: dare_ties
parameters:
  int8_mask: true
slices:
- sources:
  - layer_range: [0, 32]
    model: mistralai/Mistral-7B-v0.1
  - layer_range: [0, 32]
    model: : SanjiWatsuki/Kunoichi-DPO-v2-7B
    parameters:
      density: 0.8
      weight: 0.5
  - layer_range: [0, 32]
    model: : maywell/PiVoT-0.1-Evil-a
    parameters:
      density: 0.3
      weight: 0.15
name: first-step
---
base_model: mistralai/Mistral-7B-v0.1
dtype: float16
merge_method: dare_ties
parameters:
  int8_mask: true
slices:
- sources:
  - layer_range: [0, 32]
    model: mistralai/Mistral-7B-v0.1
  - layer_range: [0, 32]
    model: mlabonne/ArchBeagle-7B
    parameters:
      density: 0.8
      weight: 0.75
  - layer_range: [0, 32]
    model: LakoMoor/Silicon-Alice-7B
    parameters:
      density: 0.6
      weight: 0.30
name: second-step
---
models:
   - model: first-step
   - model: second-step
merge_method: slerp
base_model: first-step
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
  int8_mask: true
  normalize: true
dtype: float16

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	72.35
AI2 Reasoning Challenge (25-Shot)	69.62
HellaSwag (10-Shot)	87.14
MMLU (5-Shot)	65.11
TruthfulQA (0-shot)	61.08
Winogrande (5-shot)	81.22
GSM8k (5-shot)	69.90