xDAN-SlimOrca: A Merged 7B Language Model
xDAN-SlimOrca is a 7 billion parameter language model developed by Azazelle, created through a slerp merge of two distinct models: xDAN-L1-Chat-RL-v1 and mistral-7b-slimorcaboros. This merging technique, specified in the provided mergekit YAML configuration, allows for a nuanced combination of the strengths of its constituent models, with specific t values applied to different layers (self_attn, mlp) and a fallback for other tensors.
Key Capabilities & Performance
Based on the Mistral-7B-v0.1 architecture, xDAN-SlimOrca is designed for general-purpose language understanding and generation. Its performance has been evaluated on the Open LLM Leaderboard, achieving an overall average score of 68.04. Specific benchmark results include:
- AI2 Reasoning Challenge (25-Shot): 65.61
- HellaSwag (10-Shot): 85.70
- MMLU (5-Shot): 63.67
- TruthfulQA (0-shot): 57.68
- Winogrande (5-shot): 77.66
- GSM8k (5-shot): 57.92
These scores indicate a balanced capability across reasoning, common sense, and factual recall tasks. The model's 4096-token context length supports processing moderately long inputs and generating coherent responses.
Good For
- General conversational AI applications requiring a 7B parameter model.
- Tasks benefiting from a blend of capabilities derived from its merged base models.
- Developers seeking a Mistral-based model with specific fine-tuning characteristics from the merged components.