According to Shen et al. (2023), the series Nous-Hermes, which was developed by Nous Research and supported by Redmond AI, represents one of the most advanced generations of language models. This series, comprised of models with over 300,000 instructions, is based on Calls y Llama-2. They offer solid performance across a variety of tasks and are notable for their extensive answers, low rate of hallucinations and the absence of censorship, which makes them a competitive option in the field of Artificial Intelligence.
Training and Optimization of the Nous-Hermes Models
The models Nous-Hermes They have been perfected through rigorous machine training DGX 8x A100 80GB for long periods. The goal has been to maintain consistency with the original model Hermes and improve their capabilities. Synthetic data have been used from GPT-4 along with various sources such as GPTeacher, versions of role-playing games 1 and 2.2, code instructions, and exclusive sets such as Nous Instruct & PDACTL, as well as specific sets such as Camel-AI y Airoboros (González-Santamarta et al., 2023).
Optimization for Role-Playing Games and Hardware Recommendations
The version Llama2-13b It was designed for users who prefer to maintain consistency with the model. Hermes previous. In addition, a subset of the models Nous-Hermes has been adapted for role-playing games wearing LIMARP Lora, with a smaller but high-quality dataset. This approach differs from conventional strategies that use large numbers of examples of varying quality (González-Santamarta et al., 2023).
Shen et al. (2023) highlight that the performance of Nous-Hermes It depends on the hardware used. They provide specific recommendations for the versions. GPTQ y GGML/GGUF, taking into account the requirements of VRAM y RAM depending on the model's size. They also emphasize the importance of the RAM bandwidth and the model size to optimize the inference speed.
