Nanbeige2-16B-Chat

Introduction

The Nanbeige2-16B-Chat is the latest 16B model developed by the Nanbeige Lab, which utilized 4.5T tokens of high-quality training data during the training phase. During the alignment phase, we initially trained our model using 1 million samples through Supervised Fine-Tuning (SFT). We then engaged in curriculum learning with 400,000 high-quality samples that presented a greater level of difficulty. Subsequently, we incorporated human feedback through the Direct Preference Optimization (DPO), culminating in the development of Nanbeige2-16B-Chat. Nanbeige2-16B-Chat has achieved superior performance across various authoritative benchmark datasets.

Evaluation

We have evaluated Nanbeige2-16B-Chat's general question-answering capabilities and human preference alignments on several popular benchmark datasets. The model has achieved notable results in single-turn English QA (AlpacaEval 2.0), single-turn Chinese QA (AlignBench), and multi-turn English QA (MT-Bench).

AlpacaEval 2.0(LC Win Rate/ Win Rate)	AlignBench	MT-Bench
43.0%/40.4%	7.62	8.60

Inference

from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(
  'Nanbeige/Nanbeige2-16B-Chat',
  use_fast=False,
  trust_remote_code=True
)
model = AutoModelForCausalLM.from_pretrained(
  'Nanbeige/Nanbeige2-16B-Chat',
  torch_dtype='auto',
  device_map='auto',
  trust_remote_code=True
)
messages = [
  {'role': 'user', 'content': 'Hello'}
]
prompt = tokenizer.apply_chat_template(
  messages,
  add_generation_prompt=True,
  tokenize=False
)
input_ids = tokenizer(prompt, add_special_tokens=False, return_tensors='pt').input_ids
output_ids = model.generate(input_ids.to('cuda'))
resp = tokenizer.decode(output_ids[0][len(input_ids[0]):], skip_special_tokens=True)
print(resp)

Limitations

While we place great emphasis on the safety of the model during the training process, striving to ensure that its outputs align with ethical and legal requirements, it may not completely avoid generating unexpected outputs due to the model's size and probabilistic nature. These outputs may include harmful content such as bias or discrimination. Please don't propagate such content. We do not assume any responsibility for the consequences resulting from the dissemination of inappropriate information.

License

When using the Nanbeige models, you must comply with the Apache 2.0 License and the License Agreement for Large Language Models Nanbeige. If you intend to use the Nanbeige Models or its derivatives for commercial purposes, please submit application materials to meet the requirements of the Nanbeige Models Community License Agreement by contacting nanbeige@126.com. After review, We will grant you a non-exclusive, worldwide, non-transferable, non-sublicensable and revocable commercial copyright license.