Open_Gpt4_v0.2
This is the un-quantized fp16 version for training and merging. If you want the quantized version for inference please refer to the repo bellow:
This model is a TIES merger of Mixtral-8x7B-Instruct-v0.1 and bagel-dpo-8x7b-v0.2 with MixtralOrochi8x7B being the Base model.
I was very impressed with MixtralOrochi8x7B performance and multifaceted usecases as it is already a merger of many usefull Mixtral models such as Mixtral instruct, Noromaid-v0.1-mixtral, openbuddy-mixtral and possibly other models that were not named. My goal was to expand the models capabilities and make it even more useful of a model, maybe even competitive with closed source models like Gpt-4. But for that more testing is required. I hope the community can help me determine if its deserving of its name. 😊
This is the second iteration of this model, using better models in the merger to improve performance (hopefully).
Base model:
Merged models:
Instruct template: Alpaca
Merger config:
models:
- model: Mixtral-8x7B-Instruct-v0.1
parameters:
density: .5
weight: 1
- model: bagel-dpo-8x7b-v0.2
parameters:
density: .5
weight: .7
merge_method: ties
base_model: MixtralOrochi8x7B
parameters:
normalize: true
int8_mask: true
dtype: float16
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 73.59 |
| AI2 Reasoning Challenge (25-Shot) | 68.69 |
| HellaSwag (10-Shot) | 86.16 |
| MMLU (5-Shot) | 72.07 |
| TruthfulQA (0-shot) | 71.92 |
| Winogrande (5-shot) | 83.58 |
| GSM8k (5-shot) | 59.14 |