TinyLlama-Hybrid-Merge

This is a merge of TinyLlama models created using MergeKit, combining the foundational capabilities of the base TinyLlama with its Chat-tuned version through a sophisticated SLERP fusion with variable interpolation values.

About Me

I'm David Soeiro-Vuong, a third-year Computer Science student working as an apprentice at TW3 Partners, a company specialized in Generative AI. Passionate about artificial intelligence and language models optimization, I focus on creating efficient model merges that balance performance and capabilities.

🔗 Connect with me on LinkedIn

Merge Details

Merge Method

This model uses SLERP (Spherical Linear Interpolation) with carefully tuned parameters to achieve optimal performance balance:

• Attention Layers: Variable interpolation values [0, 0.5, 0.3, 0.7, 1] leveraging the chat model's instruction-following capabilities
• MLP Layers: Variable interpolation values [1, 0.5, 0.7, 0.3, 0] maintaining the base model's reasoning capabilities
• Other Parameters: 0.5 interpolation value creating an equal blend for balanced performance
• Format: bfloat16 precision for efficient memory usage

Models Merged

• TinyLlama/TinyLlama-1.1B-step-50K-105b - The base TinyLlama model offering foundational language capabilities
• TinyLlama/TinyLlama-1.1B-Chat-v1.0 - A fine-tuned version optimized for chat and instruction following

Configuration

slices:
  - sources:
      - model: TinyLlama/TinyLlama-1.1B-step-50K-105b
        layer_range: [0, 22]
      - model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
        layer_range: [0, 22]
merge_method: slerp
base_model: TinyLlama/TinyLlama-1.1B-step-50K-105b
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: bfloat16

Model Capabilities

This merge combines:

• TinyLlama base model's foundational knowledge and reasoning
• TinyLlama Chat's improved instruction following and conversational abilities
• Optimized parameter distribution for balanced performance
• Compact 1.1B parameter size suitable for resource-constrained environments

The resulting model provides enhanced performance on tasks requiring both reasoning and conversational abilities, such as:

• Basic question answering with improved coherence
• Simple instruction following with better response quality
• Lightweight deployment scenarios requiring balanced capabilities
• Educational and demonstration purposes for model merging techniques

Limitations

• Inherits the fundamental limitations of small 1.1B parameter models
• Limited context window and knowledge compared to larger models
• May struggle with complex reasoning, specialized domains, or nuanced tasks
• No additional training beyond the parameter merging process
• Performance ceiling constrained by the small model size

License

This model is released under the Apache 2.0 license, consistent with the underlying models' licenses.