Mistral-LLaMA-Fusion is an experimental merged language model combining the strengths of Mistral-7B-v0.1 and LLaMA-2-7B using the Linear Merge method via MergeKit. This hybrid model aims to balance Mistral’s efficiency and architecture with LLaMA-2’s robustness in reasoning and instruction following.

🔗 Created by: [Matteo Khan]
🎓 Affiliation: Apprentice at TW3 Partners (Generative AI Research)
📍 License: MIT

🔗 Connect on LinkedIn
🔗 Model on Hugging Face

🧠 Model Details

Model Type: Merged Language Model
Parent Models:
- Mistral-7B-v0.1
- LLaMA-2-7B
Merging Method: Linear Merge (via MergeKit)

🎯 Intended Use

This model is suited for research in model merging and hybridization, and can be used for:

✅ Text Generation
✅ Instruction Following
✅ Creative Writing
✅ Prompt Engineering Experiments

⚠️ Limitations

As with all merged models, this fusion may inherit and combine weaknesses from both parents:

❌ Possible generation of false, biased, or inappropriate content
⚠️ Unpredictable behavior in edge cases
📉 No guaranteed performance gain across all benchmarks

🔬 Merging Configuration

yaml

merge_method: linear
dtype: float16
models:
  - model: mistralai/Mistral-7B-v0.1
    parameters:
      t: 1.0
      weight: 0.6
  - model: meta-llama/Llama-2-7b-hf
    parameters:
      t: 1.0
      weight: 0.4

parameters:
  normalize: true
  int8_mask: false

layers:
  - pattern: "model.*"
📌 Note: No additional fine-tuning was performed. This is a straight merge using MergeKit.

🌱 Why Merging?
Merging allows rapid experimentation with existing checkpoints while reducing the computational cost and carbon footprint compared to training from scratch.

🚀 How to Use
python
Copier
Modifier
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "MatteoKhan/Mistral-LLaMA-Fusion"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto")

prompt = "Explain the benefits of merging language models."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))