Tiny-llama

Model Description

Tiny llamix is a model built from TinyLlama using Charles Goddard's mergekit on the mixtral branch. Though techincally a mixtral model it can be plugged into most llama implementation (Maybe...). The model uses Tiny-llama's tokenizer and works on the same prompt format.

This model is a proof-of-concept and might not yield necessarily better outputs. (IDK haven't tested it...)

Configuration

base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
gate_mode: hidden
dtype: bfloat16 
experts:
  - source_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
    positive_prompts:
      - "M1"
  - source_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
    positive_prompts:
     - "M2"

Usage

It can be used like any other model

from transformers import AutoModelForCausalLM, AutoTokenizer
#load model and tokenizer
model = AutoModelForCausalLM.from_pretrained("SE6446/Tiny-llamix").to("cuda")
tokenizer = AutoTokenizer.from_pretrained("SE6446/Tiny-llamix")
#write and tokenize prompt
instruction = '''<|system|>\nYou are a chatbot who can help code!</s>
<|user|> Write me a function to calculate the first 10 digits of the fibonacci sequence in Python and print it out to the CLI.</s>
<|assistant|>'''
inputs = tokenizer(instruction, return_tensors="pt", return_attention_mask=False).to("cuda")

#generate
outputs = model.generate(**inputs, max_length=200)

#print
text = tokenizer.batch_decode(outputs)[0]
print(text)

Acknowledgements

To Charles Goddard for creating the tool and for explaining it in his blog in a way a buffoon like me could understand.

To TinyLlama for providing the model as open source!