MrOvkill_-_Phi-3-Instruct-Bloated-gguf

Quantization made by Richard Erkhov.

Github

Discord

Request more models

Phi-3-Instruct-Bloated - GGUF

• Model creator: https://huggingface.co/MrOvkill/
• Original model: https://huggingface.co/MrOvkill/Phi-3-Instruct-Bloated/

Original model description:

tags:

• merge
• mergekit
• lazymergekit
• microsoft/Phi-3-mini-128k-instruct
• NexaAIDev/Octopus-v4 base_model:
• microsoft/Phi-3-mini-128k-instruct
• NexaAIDev/Octopus-v4 license: mit language:
• en library_name: transformers pipeline_tag: text-generation

---

Phi-3-Instruct-Bloated

Phi-3-Instruct-Bloated is a merge of the following models using LazyMergekit:

• microsoft/Phi-3-mini-128k-instruct
• NexaAIDev/Octopus-v4

🧩 Configuration

slices:
  - sources:
      - model: microsoft/Phi-3-mini-128k-instruct
        layer_range: [0, 32]
      - model: NexaAIDev/Octopus-v4
        layer_range: [0, 32]
merge_method: slerp
base_model: microsoft/Phi-3-mini-128k-instruct
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

# Installation
!pip install -qU transformers accelerate

# Imports
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Loading
tokenizer = AutoTokenizer.from_pretrained("MrOvkill/Phi-3-Instruct-Bloated")
model = AutoModelForCausalLM.from_pretrained("MrOvkill/Phi-3-Instruct-Bloated")

# Completion function
def infer(prompt, **kwargs):
    inputs = tokenizer(prompt, return_tensors="pt")
    with torch.no_grad():
        outputs = model.generate(**inputs, **kwargs)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Some silliness
infer("<|user|>\nBen is going to the store for some Ice Cream. So is Jerry. They mix up the ice cream at the store. Is the ice cream: (a. Ben's (b. Jerry's (c. Ben and Jerry's <|end|>\n<|assistant|>\nMy answer is (", max_new_tokens=1024)

# A proper test
infer(
    """
<|user|>
Explain what a Mixture of Experts is in less than 100 words.
<|assistant|>
""",
    max_new_tokens=1024,
    do_sample=False,
    temperature=0.0,
    top_k=50,
    top_p=0.89,
)