Pearl-7B-slerp, an xtraordinary 7B model for maths

03-22-2024 - To date, louisbrulenaudet/Pearl-34B-ties is the "Best 🤝 base merges and moerges model of around 30B" on the Open LLM Leaderboard.

Pearl-7B-slerp is a merge of the following models:

• mlabonne/OmniBeagle-7B
• WizardLM/WizardMath-7B-V1.1

Evaluation

The evaluation was performed using the HuggingFace Open LLM Leaderboard.

Model	Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K	Params (B)
louisbrulenaudet/Pearl-7B-slerp	72.75	68.00	87.16	64.04	62.35	81.29	73.62	7.24
mistralai/Mixtral-8x7B-Instruct-v0.1	72.62	70.22	87.63	71.16	64.58	81.37	60.73	46.7
microsoft/phi-2	61.33	61.09	75.11	58.11	44.47	74.35	54.81	2.78
microsoft/Orca-2-13b	58.64	60.67	79.81	60.37	56.41	76.64	17.97	13
mistralai/Mistral-7B-Instruct-v0.1	54.96	54.52	75.63	55.38	56.28	73.72	14.25	7.24
meta-llama/Llama-2-7b-hf	50.97	53.07	78.59	46.87	38.76	74.03	14.48	6.74

Spherical Linear Interpolation (SLERP) serves as a technique for seamlessly interpolating between two vectors while maintaining a constant rate of change and upholding the geometric properties of the spherical space in which these vectors exist.

Opting for SLERP over traditional linear interpolation is motivated by various considerations. Linear interpolation in high-dimensional spaces may result in a reduction in the magnitude of the interpolated vector, diminishing the scale of weights. Additionally, in many cases, the alteration in the weights' direction conveys more meaningful information, such as feature learning and representation, compared to the magnitude of change.

$${\displaystyle \mathrm{slerp}(p_0,p_1;t)=\frac{\sin [(1-t)\mathrm{\Omega }]}{\sin \mathrm{\Omega }}{p}_0+\frac{\sin [t\mathrm{\Omega }]}{\sin \mathrm{\Omega }}{p}_1\mathrm{.}}\{\backslash displaystyle\; \backslash operatorname\; \{slerp\}\; (p\_\{0\},p\_\{1\};t)=\{\backslash frac\; \{\backslash sin\; \{[(1-t)\backslash Omega\; \}]\}\{\backslash sin\; \backslash Omega\; \}\}p\_\{0\}+\{\backslash frac\; \{\backslash sin[t\backslash Omega\; ]\}\{\backslash sin\; \backslash Omega\; \}\}p\_\{1\}.\}$$

The implementation of SLERP involves the following steps:

• Normalize the input vectors to unit length, ensuring they signify directions rather than magnitudes.
• Calculate the angle between these vectors using their dot product.
• If the vectors are nearly collinear, the method defaults to linear interpolation for efficiency. Otherwise, SLERP calculates scale factors based on the interpolation factor t (where t=0 corresponds to 100% of the first vector, and t=1 corresponds to 100% of the second vector) and the angle between the vectors.
• Utilize these computed factors to weigh the original vectors, and then sum them to derive the interpolated vector.

In essence, SLERP provides a robust mechanism for interpolating vectors, offering advantages in preserving directional information and mitigating issues associated with linear interpolation in high-dimensional spaces.

Configuration

slices:
  - sources:
      - model: mlabonne/OmniBeagle-7B
        layer_range: [0, 32]
      - model: WizardLM/WizardMath-7B-V1.1
        layer_range: [0, 32]
merge_method: slerp
base_model: mlabonne/OmniBeagle-7B
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

Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "louisbrulenaudet/Pearl-7B-slerp"
messages = [{"role": "user", "content": "What is a large language model?"}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])

Citing & Authors

If you use this code in your research, please use the following BibTeX entry.

@misc{louisbrulenaudet2023,
  author =       {Louis Brulé Naudet},
  title =        {Pearl-7B-slerp, an xtraordinary 7B model for maths},
  year =         {2023}
  howpublished = {\url{https://huggingface.co/louisbrulenaudet/Pearl-7B-slerp}},
}

Feedback

If you have any feedback, please reach out at louisbrulenaudet@icloud.com.