saishf/SOVLish-Devil-8B-L3

merge

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

Merge Details

Devil >:3

This is another "SOVL" style merge, this time using mlabonne/Daredevil-8B-abliterated.

Daredevil is the first abliterated model i've tried that feels as smart as base llama-3-instruct while also being willing to give instructions to do all kinda of illegal things

This model should do well in rp, I'm yet to test it (waiting for gguf files @_@)

Merge Method

This model was merged using the Model Stock merge method using mlabonne/Daredevil-8B-abliterated as a base.

Models Merged

The following models were included in the merge:

• mlabonne/Daredevil-8B-abliterated + ResplendentAI/RP_Format_QuoteAsterisk_Llama3
• mlabonne/Daredevil-8B-abliterated + ResplendentAI/BlueMoon_Llama3
• mlabonne/Daredevil-8B-abliterated + ResplendentAI/Luna_Llama3
• mlabonne/Daredevil-8B-abliterated + ResplendentAI/Aura_Llama3
• mlabonne/Daredevil-8B-abliterated + ResplendentAI/Smarts_Llama3

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: mlabonne/Daredevil-8B-abliterated+ResplendentAI/Aura_Llama3
  - model: mlabonne/Daredevil-8B-abliterated+ResplendentAI/Smarts_Llama3
  - model: mlabonne/Daredevil-8B-abliterated+ResplendentAI/Luna_Llama3
  - model: mlabonne/Daredevil-8B-abliterated+ResplendentAI/BlueMoon_Llama3
  - model: mlabonne/Daredevil-8B-abliterated+ResplendentAI/RP_Format_QuoteAsterisk_Llama3
merge_method: model_stock
base_model: mlabonne/Daredevil-8B-abliterated
dtype: bfloat16

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	71.86
AI2 Reasoning Challenge (25-Shot)	69.20
HellaSwag (10-Shot)	84.44
MMLU (5-Shot)	68.97
TruthfulQA (0-shot)	57.95
Winogrande (5-shot)	78.14
GSM8k (5-shot)	72.48