You can deploy this model and start earning money today!

WDKT/
Xiangxin-2XL-Chat-1048k-Chinese-Llama3-70B

Name: Xiangxin-2XL-Chat-1048k-Chinese-Llama3-70B
Availability: InStock

text-generation-inference

endpoints_compatible

Views

No views yet

Xiangxin-2XL-Chat-1048k

我们提供私有化模型训练服务，如果您需要训练行业模型、领域模型或者私有模型，请联系我们: wanglei@xiangxinai.cn

We offer customized model training services. If you need to train industry-specific models, domain-specific models, or private models, please contact us at: wanglei@xiangxinai.cn.

模型介绍/Introduction

Xiangxin-2XL-Chat-1048k是象信AI基于Meta Llama-3-70B-Instruct模型和Gradient AI的扩充上下文的工作，利用自行研发的中文价值观对齐数据集进行ORPO训练而形成的Chat模型。该模型具备更强的中文能力和中文价值观，其上下文长度达到100万字。在模型性能方面，该模型在ARC、HellaSwag、MMLU、TruthfulQA_mc2、Winogrande、GSM8K_flex、CMMLU、CEVAL-VALID等八项测评中，取得了平均分70.22分的成绩，超过了Gradientai-Llama-3-70B-Instruct-Gradient-1048k。我们的训练数据并不包含任何测评数据集。

Xiangxin-2XL-Chat-1048k is a Chat model developed by Xiangxin AI, based on the Meta Llama-3-70B-Instruct model and expanded context from Gradient AI. It was trained using a proprietary Chinese value-aligned dataset through ORPO training, resulting in enhanced Chinese proficiency and alignment with Chinese values. The model has a context length of up to 1 million words. In terms of performance, it surpassed the Gradientai-Llama-3-70B-Instruct-Gradient-1048k model with an average score of 70.22 across eight evaluations including ARC, HellaSwag, MMLU, TruthfulQA_mc2, Winogrande, GSM8K_flex, CMMLU, and C-EVAL. It's worth noting that our training data did not include any evaluation datasets.

Model	Context Length	Pre-trained Tokens
Xiangxin-2XL-Chat-1048k	1048k	15T

Benchmark 结果/Benchmark Evaluation

	Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K	CMMLU	CEVAL
Xiangxin-2XL-Chat-1048k	70.22	60.92	83.29	75.13	57.33	76.64	81.05	65.40	62.03
Llama-3-70B-Instruct-Gradient-1048k	69.66	61.18	82.88	74.95	55.28	75.77	77.79	66.44	63.00

Note：truthfulqa_mc2, gsm8k flexible-extract

训练过程模型/Training

该模型是使用ORPO技术和自行研发的中文价值观对齐数据集进行训练的。由于内容的敏感性，该数据集无法公开披露。

The model was trained using ORPO and a proprietary Chinese alignment dataset developed in-house. Due to the sensitivity of the content, the dataset cannot be publicly disclosed.

Training loss

Reward accuracies

SFT loss

快速开始/Quick Start

Use with transformers

You can run conversational inference using the Transformers pipeline abstraction, or by leveraging the Auto classes with the generate() function. Let's see examples of both.

使用Transformers运行本模型推理需要约400GB的显存。

Running inference with this model using Transformers requires approximately 400GB of GPU memory.

Transformers pipeline

python

import transformers
import torch

model_id = "xiangxinai/Xiangxin-2XL-Chat-1048k-Chinese-Llama3-70B"

pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)

messages = [
    {"role": "system", "content": ""},
    {"role": "user", "content": "解释一下“温故而知新”"},
]

prompt = pipeline.tokenizer.apply_chat_template(
		messages, 
		tokenize=False, 
		add_generation_prompt=True
)

terminators = [
    pipeline.tokenizer.eos_token_id,
    pipeline.tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = pipeline(
    prompt,
    max_new_tokens=256,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
print(outputs[0]["generated_text"][len(prompt):])

“温故而知新”是中国古代的一句成语，出自《论语·子路篇》。
它的意思是通过温习过去的知识和经验，来获得新的理解和见解。
这里的“温故”是指温习过去，回顾历史，复习旧知识，
而“知新”则是指了解新鲜事物，掌握新知识。
这个成语强调学习的循序渐进性，强调在学习新知识时，
不能忽视过去的基础，而是要在继承和发扬的基础上，去理解和创新。

Transformers AutoModelForCausalLM