Undi95/ReMM-v2-L2-13B
Undi95/ReMM-v2-L2-13B is a 13 billion parameter language model, a recreation of the original MythoMax-L2-13b, developed by Undi95. This model is created using the SLERP merging method, combining updated versions of ReML v2 and Huginn v1.2. It is designed for general language tasks, leveraging its merged architecture for improved performance across various benchmarks.
Loading preview...
Undi95/ReMM-v2-L2-13B: An Updated MythoMax Recreation
Undi95/ReMM-v2-L2-13B is a 13 billion parameter language model that serves as a recreation of the popular MythoMax-L2-13b, incorporating more recent base models. This model was developed by Undi95 using the SLERP merging method to combine two key components: an updated ReML v2 and Huginn v1.2.
Key Merging Details:
- ReML v2 itself is a merge of:
- The-Face-Of-Goonery/Chronos-Beluga-v2-13bfp16
- jondurbin/spicyboros-13b-2.2 (replacing an older Airoboros version)
- NousResearch/Nous-Hermes-Llama2-13b
- Huginn v1.2 refers to The-Face-Of-Goonery/Huginn-13b-v1.2.
Performance Benchmarks:
Evaluated on the Open LLM Leaderboard, ReMM-v2-L2-13B demonstrates competitive performance for its size:
- Avg. Score: 50.57
- ARC (25-shot): 61.95
- HellaSwag (10-shot): 84.0
- MMLU (5-shot): 56.14
- TruthfulQA (0-shot): 50.81
- Winogrande (5-shot): 75.85
Prompt Format:
The model utilizes the Alpaca prompt template for instruction-following tasks, making it compatible with common instruction-tuned workflows.
Good for:
- General-purpose text generation and understanding.
- Users seeking an updated alternative to MythoMax-L2-13b.
- Experimenting with models built on advanced merging techniques.