louisbrulenaudet/Pearl-7B-0211-ties-GGUF

Pearl-7B-0211-ties, an xtraordinary 7B model

Pearl-7B-0211-ties is a merge of the following models:

• louisbrulenaudet/Pearl-7B-slerp
• WizardLM/WizardMath-7B-V1.1
• cognitivecomputations/WestLake-7B-v2-laser
• CultriX/NeuralTrix-7B-dpo

Evaluation

The evaluation was performed using the HuggingFace Open LLM Leaderboard.

Model	Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K	#Params (B)
louisbrulenaudet/Pearl-34B-ties	75.48	70.99	84.83	76.63	70.32	82.64	67.48	34.39
louisbrulenaudet/Pearl-7B-0211-ties	75.11	71.42	88.86	63.91	71.46	84.37	70.66	7.24
NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO	73.35	71.08	87.29	72.17	54.83	83.11	71.65	46.7
argilla/notus-8x7b-experiment	73.18	70.99	87.73	71.33	65.79	81.61	61.64	46.7
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	72.7	70.14	87.55	71.4	64.98	81.06	61.11	46.7
microsoft/Orca-2-13b	61.98	60.92	79.85	60.3	56.42	76.56	37.83	13
microsoft/phi-2	61.33	61.09	75.11	58.11	44.47	74.35	54.81	2.78

Ties merging

TIES-Merging is a method designed to facilitate the efficient merging of multiple task-specific models into a consolidated multitask model. It addresses two primary challenges encountered in the process of model merging with a focus on maintaining objectivity.

One key challenge tackled by TIES-Merging involves addressing redundancy in model parameters. This is achieved by identifying and eliminating redundant parameters within task-specific models, emphasizing the changes made during fine-tuning and selectively retaining the top-k% most significant changes while discarding the rest.

Another challenge pertains to conflicts arising from disagreements between parameter signs across different models. TIES-Merging resolves these conflicts by creating a unified sign vector representing the most dominant direction of change across all models.

The TIES-Merging process consists of three steps:

• Trim: Reduces redundancy in task-specific models by retaining a fraction of the most significant parameters (density parameter) and resetting the remaining parameters to zero.
• Elect Sign: Resolves sign conflicts across different models by creating a unified sign vector based on the most dominant direction (positive or negative) in terms of cumulative magnitude.
• Disjoint Merge: Averages parameter values aligned with the unified sign vector, excluding zero values.

Configuration

models:
  - model: OpenPipe/mistral-ft-optimized-1227
  - model: louisbrulenaudet/Pearl-7B-slerp
    parameters:
      density: 0.6
      weight: 0.3
  - model: WizardLM/WizardMath-7B-V1.1
    parameters:
      density: 0.55
      weight: 0.2
  - model: cognitivecomputations/WestLake-7B-v2-laser
    parameters:
      density: 0.55
      weight: 0.25
  - model: CultriX/NeuralTrix-7B-dpo
    parameters:
     density: 0.6
     weight: 0.25
merge_method: ties
base_model: OpenPipe/mistral-ft-optimized-1227
parameters:
  normalize: true
  int8_mask: true
dtype: float16

Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "louisbrulenaudet/Pearl-7B-0211-ties"
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-0211-ties, an xtraordinary 7B model},
  year =         {2023}
  howpublished = {\url{https://huggingface.co/louisbrulenaudet/Pearl-7B-0211-ties}},
}

Feedback

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