<p align="left"> <a href="https://go.upstage.ai/3Xk9J6X"> </a> <p>

Solar Pro Preview: The most intelligent LLM on a single GPU

Summary

We introduce Solar Pro Preview, an advanced large language model (LLM) with 22 billion parameters designed to fit into a single GPU. Solar Pro Preview shows superior performance compared to LLMs with less than 30 billion parameters and delivers performance comparable to models over three times its size, such as Llama 3.1 with 70 billion parameters.

Solar Pro Preview is developed using an enhanced version of our previous depth up-scaling method, which scales a Phi-3-medium model with 14 billion parameters to 22 billion parameters, intended to run on a GPU with 80GB of VRAM. Our carefully curated training strategy and dataset have significantly enhanced performance from Phi-3-medium, particularly on the MMLU-Pro and IFEval benchmarks, both respected for evaluating a model’s knowledge and instruction-following abilities.

Solar Pro Preview is a pre-release version of the official Solar Pro, with limitations on language coverage and a maximum context length of 4K. However, we believe Solar Pro Preview not only stands out as a highly efficient and capable model, but has the potential to be further extended to cover more languages and capabilities. The official version of Solar Pro will be released this November 2024 with expanded language support beyond English and longer context windows. To stay informed about the latest updates, please sign up for our mailing list. If you have any feedback or questions about the model, please visit our model discussion board and connect with us directly.

Usage

Solar Pro Preview is an instruction-tuned language model. This model is specifically designed to follow instructions and engage in conversational tasks.

Chat Template

As an instruction-tuned model, Solar Pro Preview uses the ChatML template for optimal performance in conversational and instruction-following tasks. This approach aligns with the model's training data and is likely to yield more accurate and relevant responses. For instance, a question formatted in the ChatML template looks like the following, where the model generates the answer after <|im_start|>assistant. Note that system prompts are not currently supported in Solar Pro Preview. This feature will be available in the official release.

<|im_start|>user
Please, introduce yourself.<|im_end|>
<|im_start|>assistant

Text Generation

Below is an example inference code that details loading the model, applying the chat template, and generating the model answer.

# Install requirements
# !pip install transformers==4.44.2 torch==2.3.1 flash_attn==2.5.8 accelerate==0.31.0

# Load model
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("upstage/solar-pro-preview-instruct")
model = AutoModelForCausalLM.from_pretrained(
    "upstage/solar-pro-preview-instruct",
    device_map="cuda",  
    torch_dtype="auto",  
    trust_remote_code=True,
)
# Apply chat template
messages = [
    {"role": "user", "content": "Please, introduce yourself."},
]
prompt = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
# Generate text
outputs = model.generate(prompt, max_new_tokens=512)
print(tokenizer.decode(outputs[0]))

Solar Pro Preview is also available as an API in Upstage Console and we provide other easy-to-use methods as well. If you'd like to explore these options, please visit our blog page.

Evaluation

Solar Pro Preview is evaluated over a variety of benchmarks.

	Solar-pro-preview	Phi-3-medium-4K-instruct	Phi-3.5-MoE-instruct	Gemma 2 27B IT	Llama-3.1-8B-instruct	Llama-3.1-70B-instruct
Release Date	2024.09.08	2024.05.02	2024.08.20	2024.06.25	2024.06.18	2024.06.16
Model size	22B	14B	41.9B (6.6B)	27B	8B	70B
License	MIT	MIT	MIT	gemma	llama3.1	llama3.1
MMLU	79.14	78.02	78.66	76.13	68.25	82.09
MMLU Pro	52.11	47.51	46.99	45.68	37.88	53.01
IFEval	84.37	64.37	69.15	75.36	77.40	84.13
ARC-C	68.86	66.55	68.34	74.06	60.24	70.39
GPQA	36.38	35.78	34.38	36.38	35.26	41.06
HellaSwag	86.36	85.68	85.97	86.02	80.08	86.42
EQBench	77.91	76.78	77.22	80.32	65.80	82.52
BigBench Hard	67.31	63.09	62.58	64.88	51.06	69.54

...