OpenLlama-Stable-7B

This is a merge of pre-trained language models created using LazyMergekit, combining the foundational capabilities of OpenLM's Open Llama with StabilityAI's StableBeluga through an efficient SLERP fusion.

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: 0.7 interpolation value favoring StableBeluga's strong instruction-following capabilities
• MLP Layers: 0.5 interpolation value creating an equal blend for balanced reasoning
• Other Parameters: 0.6 interpolation value slightly favoring StableBeluga's refinements
• Format: bfloat16 precision for efficient memory usage

Models Merged

• openlm-research/open_llama_7b - An open-source reproduction of Meta's LLaMA that offers strong base capabilities
• stabilityai/StableBeluga-7B - StabilityAI's instruction-tuned variant offering improved instruction following and coherence

Configuration

slices:
  - sources:
      - model: openlm-research/open_llama_7b
        layer_range: [0, 32]
      - model: stabilityai/StableBeluga-7B
        layer_range: [0, 32]
merge_method: slerp
base_model: openlm-research/open_llama_7b
parameters:
  t:
    # Couches d'attention: préférence pour StableBeluga (0.7)
    - filter: self_attn
      value: 0.7
    # Couches MLP: équilibrées
    - filter: mlp
      value: 0.5
    # Tout le reste
    - value: 0.6
dtype: bfloat16

Model Capabilities

This merge combines:

• Open Llama's strong foundational knowledge and reasoning
• StableBeluga's improved instruction following and coherence
• Fully open architecture with no usage restrictions

The resulting model provides enhanced performance on tasks requiring both strong reasoning and good instruction following, such as:

• Detailed explanations of complex concepts
• Creative writing with coherent structure
• Problem-solving with step-by-step reasoning
• Balanced factual responses with nuanced perspectives

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "david-sv/OpenLlama-Stable-7B"  # Replace with your actual HF username
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

# For chat completions
prompt = """<human>: Explain the concept of spherical linear interpolation (SLERP) and why it's useful for merging language models.

<assistant>:"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(
    inputs["input_ids"],
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    repetition_penalty=1.1
)

print(tokenizer.decode(output[0], skip_special_tokens=True))

Limitations

• Inherits limitations from both base models
• May exhibit inconsistent behavior for certain complex reasoning tasks
• No additional alignment or fine-tuning beyond the base models' training
• Model was created through parameter merging without additional training data

License

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