🌟 Buying me coffee is a direct way to show support for this project.

Phizzle

Phizzle is a merge of the following models using LazyMergekit:

• rhysjones/phi-2-orange
• cognitivecomputations/dolphin-2_6-phi-2
• mrm8488/phi-2-coder

🧩 Configuration

models:
  - model: rhysjones/phi-2-orange
    parameters:
      density: 0.5
      weight: 0.3 
  - model: cognitivecomputations/dolphin-2_6-phi-2
    parameters:
      density: 0.5
      weight: 0.3
  - model: mrm8488/phi-2-coder
    parameters:
      density: 0.5
      weight: 0.3

merge_method: ties
base_model: rhysjones/phi-2-orange
parameters:
  normalize: true
dtype: float16

💻 Usage

!pip install -qU transformers accelerate einops

from transformers import AutoTokenizer
import transformers
import torch

model = "Isotonic/phizzle"
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",
    trust_remote_code=True
)

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"])

Evaluations

Evaluations done using mlabonne's usefull Colab notebook llm-autoeval. Also check out the alternative leaderboard at Yet_Another_LLM_Leaderboard

phizzle - Yet to be benchmarked

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
phi-2-orange	33.37	71.33	49.87	37.3	47.97
phi-2-dpo	30.39	71.68	50.75	34.9	46.93
dolphin-2_6-phi-2	33.12	69.85	47.39	37.2	46.89
phi-2	27.98	70.8	44.43	35.21	44.61