nbeerbower/Mistral-Nemo-Moderne-12B-FFT-experimental
nbeerbower/Mistral-Nemo-Moderne-12B-FFT-experimental is a 12 billion parameter experimental language model developed by nbeerbower, based on the Mahou-1.5-mistral-nemo-12B-lorablated architecture. It was fine-tuned using ORPO on the gutenberg2-dpo and gutenberg-moderne-dpo datasets. This model is noted for its erratic behavior and poor performance, making it unsuitable for production use and primarily serving as a public record of an experimental training process.
Loading preview...
Overview
nbeerbower/Mistral-Nemo-Moderne-12B-FFT-experimental is an experimental 12 billion parameter language model developed by nbeerbower. It is a full fine-tune of the Mahou-1.5-mistral-nemo-12B-lorablated model, utilizing the ORPO tuning method. The training involved 1.5 epochs on the gutenberg2-dpo and gutenberg-moderne-dpo datasets.
Key Characteristics
- Experimental Nature: This model is explicitly shared as an experimental iteration, reflecting a learning process rather than a production-ready release.
- Performance: The model exhibits erratic behavior and poor performance, making it unsuitable for practical applications.
- Training Method: Fine-tuned using the ORPO method over 1.5 epochs with 8x A100 GPUs.
Good for
- Research and Learning: Ideal for those interested in observing experimental model development, particularly the impact of ORPO tuning and potential conflicts with Mistral Instruct and ChatML formats.
- Understanding Limitations: Useful for studying how specific training parameters or data choices can lead to suboptimal model performance.
Note: This model is not recommended for any use cases requiring stable or high-quality output.