Overview

Artples/L-MChat-7b is a 7 billion parameter language model developed by Artples, created through a strategic merge of two prominent models: Nexusflow/Starling-LM-7B-beta and FuseAI/FuseChat-7B-VaRM. This model utilizes a slerp merge method, combining layers from both base models to achieve a balanced and robust performance profile. The merge configuration specifies distinct t values for self_attn and mlp layers, indicating a fine-tuned integration of their respective architectures.

Key Capabilities & Performance

L-MChat-7b demonstrates solid performance across a range of benchmarks, as evaluated on the Open LLM Leaderboard. Its average score is 69.57, with notable results including:

• HellaSwag (10-Shot): 84.59
• Winogrande (5-shot): 81.37
• AI2 Reasoning Challenge (25-Shot): 65.61
• MMLU (5-Shot): 65.44

Further evaluations on the Open LLM Leaderboard show an average of 21.02, with specific scores like:

• IFEval (0-Shot): 52.97
• BBH (3-Shot): 24.20

Usage & Licensing

Developers can easily integrate L-MChat-7b using the transformers library, with provided Python code for text generation. The model is released under the Apache 2.0 license, with a specific restriction against direct competition with OpenAI. The merge process itself was facilitated using LazyMergekit.