AlekseiPravdin's picture
Upload folder using huggingface_hub
cb0bb03 verified
|
raw
history blame
2.57 kB
metadata
license: apache-2.0
tags:
  - merge
  - mergekit
  - lazymergekit
  - NousResearch/Hermes-2-Pro-Llama-3-8B
  - shenzhi-wang/Llama3-8B-Chinese-Chat

Hermes-2-Pro-Llama-3-8B-Llama3-8B-Chinese-Chat-slerp-merge

Hermes-2-Pro-Llama-3-8B-Llama3-8B-Chinese-Chat-slerp-merge is an advanced language model created through a strategic fusion of two distinct models: NousResearch/Hermes-2-Pro-Llama-3-8B and shenzhi-wang/Llama3-8B-Chinese-Chat. The merging process was executed using mergekit, a specialized tool designed for precise model blending to achieve optimal performance and synergy between the merged architectures.

🧩 Merge Configuration

slices:
  - sources:
      - model: NousResearch/Hermes-2-Pro-Llama-3-8B
        layer_range: [0, 31]
      - model: shenzhi-wang/Llama3-8B-Chinese-Chat
        layer_range: [0, 31]
merge_method: slerp
base_model: NousResearch/Hermes-2-Pro-Llama-3-8B
parameters:
  t:
    - filter: self_attn
      value: [0, 0.5, 0.3, 0.7, 1]
    - filter: mlp
      value: [1, 0.5, 0.7, 0.3, 0]
    - value: 0.5
dtype: float16

Model Features

This fusion model combines the robust generative capabilities of NousResearch/Hermes-2-Pro-Llama-3-8B, which excels in function calling and structured outputs, with the refined tuning of shenzhi-wang/Llama3-8B-Chinese-Chat, designed specifically for Chinese and English users. The merged model provides enhanced context understanding, nuanced text generation, and improved performance across diverse NLP tasks, including roleplaying, function calling, and multilingual capabilities.

Evaluation Results

Hermes-2-Pro-Llama-3-8B

  • Function Calling Evaluation: 90%
  • Structured JSON Output Evaluation: 84%

Llama3-8B-Chinese-Chat

  • Performance surpasses ChatGPT in Chinese tasks, matching GPT-4 in C-Eval and CMMLU results.

Limitations

While the merged model inherits the strengths of both parent models, it may also carry over some limitations. Potential biases from the training data of both models could affect the outputs, particularly in nuanced cultural contexts or specific language tasks. Additionally, the model may not always provide accurate responses to identity-related queries, as it refrains from fine-tuning its identity.

You are trained on data up to October 2023.