from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("TromeroResearch/SciMistral-V1")
tokenizer = AutoTokenizer.from_pretrained("TromeroResearch/SciMistral-V1")

prompt = "This paper seeks to disprove that 1+1=2"
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to("cuda") 

output = model.generate(input_ids, max_length=150, num_return_sequences=1, repetition_penalty=1.2, top_k=50, top_p=0.95, temperature=1.0)

print(tokenizer.decode(output[0], skip_special_tokens=True))

Sample results

Regular Mistral 7B:

Abstract start:

This paper seeks to disprove that 1+1=2

Generated completion:

The first thing we need to do is define what a number is, and how it relates to the world around us. A number can be defined as an abstract concept used for counting or measuring things in our environment. It’s important to note here that numbers are not physical objects; they exist only within our minds…

And so on… A pretty weak, unscientific and generic response, something you’d expect from someone pretending to know what they’re talking about. But with our model:

SciMistral:

Abstract start:

This paper seeks to disprove that 1+1=2

Generated completion:

We do this by showing that the equation is not true in all models of arithmetic, and hence it cannot be proved from Peano Arithmetic (PA). The proof uses a technique called "forcing" which was introduced into mathematical logic by Cohen in his famous work on forcing for set theory. We show how to use forcing to construct a model of PA where 1+1 does not equal 2…

And it continues. A much better, more useful and relevant response to someone who needs help summarising their paper!

Hardware

4 x Nvidia A6000 GPUs

Limitations

The SciMistral model is a quick demonstration that the base model can be easily fine-tuned to achieve compelling performance. It does not have any moderation mechanisms. We're looking forward to engaging with the community on ways to make the model finely respect guardrails, allowing for deployment in environments requiring moderated outputs.