Magic-Dolphin-7b

The follow-up to this model has been released, check out the updated benchmarks here for Excalibur-7b

A full suite of GGUF quantizations can be found here, courtesy of RichardErkhov

A linear merge of:

• cognitivecomputations/dolphin-2.6-mistral-7b-dpo-laser
• Locutusque/Hyperion-1.5-Mistral-7B
• ibm/merlinite-7b

These three models showed excellent acumen in technical topics so I wanted to see how they would behave together in a merge. Several different ratios were tested before this release, in the end a higher weighting for merlinite-7b helped smooth out some edges. This model is a test of how LAB tuning is impacted by merges with models leveraging DPO.

Benchmark Performance

Name	Avg.	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K
Magic-Dolphin-7b	67.48	65.78	85.61	64.64	58.01	79.64	51.18
dolphin-2.6-mistral-7b-dpo-laser	67.28	66.3	85.73	63.16	61.71	79.16	47.61
merlinite-7b	64	63.65	84.52	64.91	50.15	79.72	41.09
Hyperion-1.5-Mistral-7B	61.43	60.49	83.64	63.57	41.78	78.61	40.49

This was my first experiment with merging models so any feedback is greatly appreciated.

Uses Alpaca template.

Sample Question

Merge Details

Merge Method

This model was merged using the linear merge method.

Models Merged

The following models were included in the merge:

• cognitivecomputations/dolphin-2.6-mistral-7b-dpo-laser
• Locutusque/Hyperion-1.5-Mistral-7B
• ibm/merlinite-7b

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: models/dolphin-2.6-mistral-7b-dpo-laser
    parameters:
      weight: 1.0
  - model: models/Hyperion-1.5-Mistral-7B
    parameters:
      weight: 0.3
  - model: models/merlinite-7b
    parameters:
      weight: 0.5
merge_method: linear
dtype: float16

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	67.48
AI2 Reasoning Challenge (25-Shot)	65.78
HellaSwag (10-Shot)	85.61
MMLU (5-Shot)	64.64
TruthfulQA (0-shot)	58.01
Winogrande (5-shot)	79.64
GSM8k (5-shot)	51.18