You can deploy this model and start earning money today!

LoneStriker/
Nucleus-1B-alpha-1-6.0bpw-h6-exl2

Name: Nucleus-1B-alpha-1-6.0bpw-h6-exl2
Availability: InStock

nampdn-ai/tiny-textbooks

mit

autotrain_compatible

text-generation-inference

endpoints_compatible

Views

No views yet

Nuclues 1B Alpha1

What is Nucleus?

Nucleus is a small language model based on Mistral (actually, the trimmed untrained version you can find here) and trained in different steps. First, we've pretrained it on TinyStories dataset, then TinyTextBooks to make it a more specific model. This model is just a proof of concept at this point, but showed good promises in early tests. So with proper training, can be a good product over time!

Inference

Open In Colab

First you need to install transformers and accelerate libraries in order to run this model. Then, you basically have to run the following code:

python

from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
import torch

model_name_or_id = "NucleusOrg/Nucleus-1B-alpha-1"

model = AutoModelForCausalLM.from_pretrained(model_name_or_id, torch_dtype=torch.float16, device_map="cuda")
tokenizer = AutoTokenizer.from_pretrained(model_name_or_id)

prompt = "### Lesson: Python Programming 101\n### Introduction\n"

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

generation_config = GenerationConfig(
    do_sample=True,
    top_k=1,
    temperature=0.9,
    max_new_tokens=500,
    repetition_penalty=1.5,
    pad_token_id=tokenizer.eos_token_id
)

outputs = model.generate(**inputs, generation_config=generation_config)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Prompt Format: This model does not have a specific prompt format, but the best results could be achieved with a textbook type of format like:

### Chapter 1: Elon Musk and Iron Man
Elon met Tony at a Cafe in Monaco, then they had a conversation about

You also can try something like this:

Question: Who are you?
Answer:

But since the model isn't made for chat/question answering, the result won't be good enough.

Repetition Penalty: Since most of these models like to repeat themselves, just keep that number there. You can increase or decrease it based on your liking,but keep in mind that a number lower than 1 makes the model super repetitive.

Known Issues

Since we only had 420k rows of data, a lot of information are missing on this model. Since mentioned earlier in this very model card, it's a proof of concept model.
You probably may test it with coding. Let's say that the model is terrible at coding. We may release a coding optimized model as soon as possible.

Our Team

Muhammadreza Haghiri (X (formerly Twitter) - Website - Github - LinkedIn)
Mahi Mohrechi (Website - Github - LinkedIn)

Special Thanks

LMLabs for providing 1B untrained model.
Mistral Team for providing the best open source base model ever.
Sina Rashidi, who translated Alpaca dataset to Persian.
Jupyto team for providing our infrastructure.

LoneStriker/Nucleus-1B-alpha-1-6.0bpw-h6-exl2