152334H/miqu-1-70b-sf
The 152334H/miqu-1-70b-sf model is a 69 billion parameter Llama-based causal language model, dequantized from a Q5 format to F16 and transposed to PyTorch by 152334H. It is a re-implementation of miqu-1-70b, offering improved shape rotation compared to other PyTorch versions. This model demonstrates strong performance across various benchmarks, including an average of 76.59 on the Open LLM Leaderboard, making it suitable for general text generation and reasoning tasks.
Loading preview...
Model Overview
152334H/miqu-1-70b-sf is a 69 billion parameter Llama-based causal language model, developed by 152334H. This version is a dequantized (from Q5 to F16) and PyTorch-transposed implementation of the original miqu-1-70b model, featuring corrected shape rotations compared to other PyTorch ports.
Key Capabilities & Performance
This model is designed for general text generation and demonstrates robust performance across a range of benchmarks. Key metrics include:
- Open LLM Leaderboard Average: 76.59
- AI2 Reasoning Challenge (25-Shot): 73.04
- HellaSwag (10-Shot): 88.61
- MMLU (5-Shot): 75.49
- Winogrande (5-Shot): 85.32
- GSM8k (5-Shot): 67.70
It also shows strong results on specific MMLU sub-categories, particularly in humanities and social sciences. The model's generation capabilities are showcased through examples of creative prose.
Usage and Licensing
The model can be loaded and used with the Hugging Face Transformers library, requiring significant GPU resources due to its size. It operates under the NOMERGE License, which explicitly prohibits merging its tensors with other models unless explicit permission is granted by the license holder. This includes forbidding the use of utilities like "mergekit" or "MergeMonster" in conjunction with this software.