Awesome

🐷SUS-Chat: Instruction tuning done right

<br><br>

News

• 2024-1-04: 🔥 cloudyu created a series of top ranked MOE based on our model

• 2023-12-09: 🔥 Tigerbot variant has been deleted, SUS-Chat-34B is now the the top-ranked LLaMA model and the top-ranked chat model.

• 2023-12-07: SUS-Chat-34B is now available on WiseModel🧠.

• 2023-12-06: Try SUS-Chat-34B chat-ui.

• 2023-12-05: SUS-Chat-34B is now available on ModelScope🤖

• 2023-12-05: SUS-Chat-34B is ranked 2nd in Open LLM leaderboard and surpassed all models under 70B.

• 2023-12-01: SUS-Chat-34B is now available on HuggingFace🤗.

Introduction

<img src="https://hackmd.io/_uploads/HJlDtzhBa.png" id="fig-sus" alt="Figure 1: DALL·E 2023-12-01 11.03.28 - An imposing, majestic wild boar combined with elements of a futuristic transformer robot. The boar itself should be intricately blended with these tra" />

SUS-Chat-34B is a 34B bilingual Chinese-English dialogue model, jointly released by the Southern University of Science and Technology and IDEA-CCNL. This model is based on 01-ai/Yi-34B and has been fine-tuned on millions of high-quality, multilingual instruction data. While maintaining the strong language capabilities of the base model, the SUS-Chat-34B model has improved the model’s response to human instructions through high-quality instruction fine-tuning and excels at imitating human thought processes through chains of thought. It introduces inter-instruction attention sharing in long texts, expanding the window size from 4K to 8K, significantly enhancing the usability of multi-turn dialogues.

It has surpassed all models of the same size in almost all benchmark tests and is better suited to meet the practical needs of complex multilingual tasks. Compared to larger models, SUS-Chat-34B remains highly competitive and has achieved state-of-the-art performance in our comprehensive evaluations.

SUS-Chat-34B model has the following highlights:

1. Large-scale complex instruction following data: Trained with 1.4 billion tokens of high-quality complex instruction data, covering Chinese and English, multi-turn dialogues, mathematics, reasoning, and various other types of instruction data;
2. Strong performance in general tasks: The SUS-Chat-34B model excels in numerous mainstream Chinese and English tasks, surpassing other open-source instruction fine-tuned models of the same parameter scale. It also competes well against models with larger parameter scales;
3. Longer context window and excellent multi-turn dialogue capabilities: Currently, SUS-Chat-34B supports an 8K context window, and is trained with a large amount of multi-turn instruction and single-multi-turn mixed data, demonstrating remarkable capabilities in long-text dialogue information focus and instruction follow-up.

SUS-Chat powerfully demonstrates that through the right instruction fine-tuning, academic institutions can achieve better performance without increasing model parameters, using open-source datasets and models. This bridges the gap between academia and industry in large language models and opens new possibilities for collaboration between academic and industrial sectors.

Performance

To better evaluate the performance of the SUS-Chat-34B model, we conducted assessments across multiple benchmark tests and have open-sourced the evaluation framework TLEM to facilitate replication and comparison by other researchers.

In TLEM, we utilized various benchmark tests including MMLU, CMMLU, C-Eval, BBH, GSM-8K, and MATH, to measure the model’s knowledge and thinking capabilities. In these metrics, the SUS-Chat-34B model achieved state-of-the-art performance. Additionally, we incorporated lm-eval to test SUS-Chat and similar models on winogrande, hellaswag, arc, and truthful-qa, assessing the model’s common-sense reasoning ability and susceptibility to illusions.

Overall, the SUS-Chat-34B model significantly outperformed models of similar scale and achieved the most advanced comprehensive performance.

<img src="https://github.com/SUSTech-IDEA/SUS-Chat/raw/main/assets/radar.png" id="fig-bench" alt="Figure 2: Benchmark" />

Math & Reasoning

More Tasks

Overall

To reproduce the results, please start a corresponding vllm server and refer to here.

Usage

SUS-Chat-34B is a standard LLaMA model and should be seamlessly compatible with the LLaMA ecosystem. We provide the following example to demonstrate how it can be used for multi-turn dialogues.

Feel free to open an issue if you have any questions.

from transformers import AutoModelForCausalLM, AutoTokenizer # 🤗 Transformers, or 
# from modelscope import AutoModelForCausalLM, AutoTokenizer # 🤖 ModelScope

def chat_template(messages):
    history = ""
    for message in messages:
        match message:
            case {"role": "user", "content": message}:
                history += f"### Human: {message}\n\n### Assistant: "
            case {"role": "assistant", "content": message}:
                history += message
    return history


model_path = "SUSTech/SUS-Chat-34B"
# model_path = "SUSTC/SUS-Chat-34B" # ModelScope

tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
model = AutoModelForCausalLM.from_pretrained(
    model_path, device_map="auto", torch_dtype="auto"
).eval()

messages = [{"role": "user", "content": "hi"}]

input_ids = tokenizer.encode(
    chat_template(messages), return_tensors="pt", add_special_tokens=False
).to("cuda")
output_ids = model.generate(input_ids.to("cuda"), max_length=256)
response = tokenizer.decode(
    output_ids[0][input_ids.shape[1] :], skip_special_tokens=False
)

messages.append({"role": "assistant", "content": response})

# Second round

messages.append({"role": "user", "content": "What is the capital of China?"})

input_ids = tokenizer.encode(
    chat_template(messages), return_tensors="pt", add_special_tokens=False
).to("cuda")
output_ids = model.generate(input_ids.to("cuda"), max_length=256)
response = tokenizer.decode(
    output_ids[0][input_ids.shape[1] :], skip_special_tokens=False
)

messages.append({"role": "assistant", "content": response})

Limitations

SUS-Chat has only undergone supervised fine-tuning and has not yet been trained on human preference learning. As a result, it may produce unreasonable responses in some situations and exacerbate existing issues in language models, including hallucinations, non-determinism, and cumulative errors. To achieve better performance for downstream tasks, we recommend adjusting the generation configuration parameters accordingly.

Disclaimer

During the training process, we used data compliance check algorithms to ensure the compliance of the training model as much as possible. Due to the complexity of the data and the diverse use cases of language models, we cannot guarantee that the model will produce correct and reasonable outputs in all scenarios. Please be aware that there is still a risk of the model generating problematic outputs. We will not be responsible for any risks or issues arising from misuse, misguidance, illegal use, and related misinformation, as well as data security issues related to the model.

License

This model is developed entirely for academic research and free commercial use, but it must adhere to the license from 01-ai.