ZeroXClem's picture
Update README.md
e7bfede verified
|
raw
history blame
4.79 kB
metadata
license: apache-2.0
tags:
  - merge
  - mergekit
  - lazzymergekit
  - Qwen2
  - Coder
  - Math
  - Bunnycore
  - Instruct
  - OpenBookQA
  - instruction-following
  - long-form-generation
base_model:
  - unsloth/Qwen2.5-Coder-1.5B-Instruct

ZeroXClem/Qwen2.5-1.5B-Instruct-Coder-Math-Bunnycore-Fusion

ZeroXClem/Qwen2.5-1.5B-Instruct-Coder-Math-Bunnycore-Fusion is a cutting-edge merged model that combines the finest features from instruction-following, coding, mathematical reasoning, and factual question-answering. This powerhouse is designed for high performance in diverse technical, creative, and interactive tasks.

🌟 Family Tree

This model is the fusion of the following:

These models have been seamlessly blended to create a versatile AI that excels across multiple domains.


🧬 Detailed Model Lineage

A: cyixiao/qwen-1.5B-openbookqa

  • Focuses on factual knowledge and reasoning from the OpenBookQA dataset, providing strong question-answering capabilities.

B: unsloth/Qwen2.5-Coder-1.5B-Instruct

  • Tailored for coding and instruction-following, this model enhances the ability to generate code and follow precise instructions with ease.

C: Qwen/Qwen2.5-Math-1.5B-Instruct

  • This model specializes in mathematical reasoning and logical problem-solving, making it perfect for structured tasks that require high-level thinking.

D: bunnycore/Qwen2.5-1.5B-Matrix

  • A multi-purpose model that blends instruction, math, and coding, providing a well-rounded performance in both structured and creative tasks.

E: Syed-Hasan-8503/Qwen2.5-1.5B-Instruct-WO-Adam-mini

  • Fine-tuned on conversational and identity-specific tasks, this model contributes to the model’s ability to handle conversation-heavy tasks with clarity.

F: Goekdeniz-Guelmez/Josiefied-Qwen2.5-1.5B-Instruct-abliterated-v3

  • This model brings uncensored capabilities, ensuring that the AI is flexible and adaptable in open-ended and unrestricted instruction-following scenarios.

πŸ› οΈ Merge Details

The model was merged using the DELLA merge method with bfloat16 precision, ensuring high-performance across multiple task types. Here's the configuration used for the merge:

merge_method: della
dtype: bfloat16
parameters:
  epsilon: 0.1
  lambda: 1.0
  normalize: true

base_model: unsloth/Qwen2.5-Coder-1.5B-Instruct

models:
  - model: cyixiao/qwen-1.5B-openbookqa
    parameters:
      weight: 1
      density: 0.5
  - model: unsloth/Qwen2.5-Coder-1.5B-Instruct
    parameters:
      weight: 1
      density: 0.6
  - model: Qwen/Qwen2.5-Math-1.5B-Instruct
    parameters:
      weight: 1
      density: 0.55
  - model: bunnycore/Qwen2.5-1.5B-Matrix
    parameters:
      weight: 1
      density: 0.55
  - model: Syed-Hasan-8503/Qwen2.5-1.5B-Instruct-WO-Adam-mini
    parameters:
      weight: 1
      density: 0.45
  - model: Goekdeniz-Guelmez/Josiefied-Qwen2.5-1.5B-Instruct-abliterated-v3
    parameters:
      weight: 1
      density: 0.5

🎯 Key Features & Capabilities

1. Coding and Instruction Following:

This model excels in technical coding tasks thanks to the contributions from Qwen2.5-Coder and Matrix.

2. Mathematical Reasoning:

With Qwen2.5-Math-1.5B-Instruct, the model is perfect for solving complex mathematical problems and structured logical tasks.

3. Conversational Abilities:

Fine-tuned on conversation and identity tasks, the model handles complex dialogue and conversational exchanges with Syed-Hasan-8503.

4. Uncensored Versatility:

Thanks to Josiefied-Qwen2.5, this model can operate without restrictions, making it ideal for open-ended instruction-following.


πŸ“œ License

This model is open-sourced under the Apache-2.0 License, allowing others to use and modify it freely, as long as they give proper attribution.


πŸ’‘ Tags

  • merge
  • Qwen
  • Coder
  • Math
  • Bunnycore
  • instruction-following
  • long-form-generation