mita-v1.0-7b-2-24-2025

Formulae/MITA-V1.0-7B-2-24-2025

Overview

Formulae/MITA-V1.0-7B is the first-generation MITA model, designed as a general-purpose, uncensored model. This version serves as the foundation for future MoE (Mixture of Experts) developments.

Built using the TIES merging method, MITA-V1.0-7B combines multiple fine-tuned models to create a balanced and robust generalist model.

Merge Details

• Base Model: deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
• Merged Models:
  • HumanLLMs/Human-Like-Qwen2.5-7B-Instruct
  • jeffmeloy/Qwen2.5-7B-nerd-uncensored-v0.9
• Merge Method: TIES (Trim, Elect Sign & Merge)
• Data Type: bfloat16

What is TIES?

TIES (Trim, Elect Sign & Merge) is a model merging technique designed to preserve valuable parameters when combining multiple models. Unlike naive merging methods, TIES minimizes parameter interference by:

1. Resetting parameters that changed minimally during fine-tuning.
2. Resolving sign conflicts between different models.
3. Merging only the aligned parameters to ensure stability.

📖 Reference: TIES Paper

Capabilities

✅ General-Purpose Intelligence – Balanced across multiple tasks.
✅ Uncensored Outputs – Designed for open and unrestricted conversation.
✅ Strong Reasoning – Maintains logical coherence across different domains.

Limitations & Risks

⚠ No Task Specialization – While versatile, it does not excel in any specific domain like coding or math.
⚠ Potential Biases – As an uncensored model, users should verify outputs for accuracy and ethical considerations.

Usage Disclaimer

MITA-V1.0-7B is an experimental foundation model. Users should validate critical outputs, especially for sensitive or factual queries.

Contribute

We welcome contributions, including quantizations, fine-tuning, and further enhancements.

💡 Support Us: Buy Me a Coffee

📩 Contact: [email protected]

---

Future Development

MITA-V1.0-7B is just the beginning. Future versions will integrate MoE architectures for better scalability and specialization.

Made possible with MergeKit.