DeepSeekFakeNews-LLM-7B-Chat

Deep SEEK-FAKE-NEWS LLM

<!-- markdownlint-disable first-line-h1 --> <!-- markdownlint-disable html --> <!-- markdownlint-disable no-duplicate-header --> <div align="center"> <img src="figures/logo_v1.0.png" width="100%" height="50%" alt="DeepSeekFakeNews-LLM" /> </div> <p align="center"> <a href="https://github.com/zoltol/DeepSeekFakeNews-LLM-7B-Chat"><b>Project Link</b>👁️</a> <a href="http://faculty.neu.edu.cn/tanzhenhua/zh_CN/zdylm/228486/list/index.htm"><b>Lab Link</b>👁️</a> </p>

1. Introduction of Deep SEEK-FAKE-NEWS LLM

The reasoning model Deep SEEK-FAKE-NEWS LLM was fine-tuned on the DeepSeek-LLM-7B-Chat base model, utilizing a newly constructed fake news instruction dataset built upon the MCFEND data. Compared to other reasoning models such as DeepSeek-LLM-7B-Chat, this reasoning model not only achieves significant improvements in F1 and accuracy (Acc) metrics but also excels in reasoning analysis, providing rationales that serve as well-founded supports for its conclusions. This work provides an adapter for LLMs, contributing significantly to the advancement of Artificial General Intelligence (AGI). The adapter enhances the capabilities of LLMs, enabling it to perform more effectively in the specific domain of fake news detection.

2. Model Summary

deepseekfakenews-llm-7b-chat is a 7B parameter model initialized from deepseek-llm-7b-chat and fine-tuned on the fake news instruction dataset that we constructed based on the MCFEND dataset.

• Github Page: zoltol/DeepSeekFakeNews-LLM-7B-Chat
• Huggingface Repository: zt-ai/DeepSeekFakeNews-LLM-7B-Chat
• **Demo of Chatting With DeepSeekFakeNews-LLM:

<div align="center"> <img src="figures/demo.PNG" width="100%" alt="Chat with DeepSeekFakeNews-LLM" /> </div>

3. How to Use

Here are some examples of how to use our model.

import torch
from peft import PeftModel
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig

model_name = "zt-ai/DeepSeekFakeNews-llm-7b-chat"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)
model.generation_config = GenerationConfig.from_pretrained(model_name)
model.generation_config.pad_token_id = model.generation_config.eos_token_id
model = PeftModel.from_pretrained(model, model_name)


messages = [
    {
"role": "user",
"content":
"""假新闻的表现可以总结为以下几个方面：1. 逻辑和事实矛盾。2.断章取义和误导性信息。3.夸张标题和吸引眼球的内容。4.情绪化和极端语言。5.偏见和单一立场。请从这几个方面分析新闻的真实性（真新闻或假新闻）：
发布时间：
新闻标题：
新闻内容：
"""}
]
input_tensor = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
outputs = model.generate(input_tensor.to(model.device), max_new_tokens=100)

result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=True)
print(result)

Avoiding the use of the provided function apply_chat_template, you can also interact with our model following the sample template. Note that messages should be replaced by your input.

User: {messages[0]['content']}

Assistant: 

Note: By default (add_special_tokens=True), our tokenizer automatically adds a bos_token (<｜begin▁of▁sentence｜>) before the input text. Additionally, since the system prompt is not compatible with this version of our models, we DO NOT RECOMMEND including the system prompt in your input.

4. Dataset

The SEEK-FAKE-NEWS LLM is post-trained on the fake news instruction dataset that we constructed based on the MCFEND dataset.

5. Evaluation Results

6. Citation

@misc {tao_zhang_2025,
	author       = { {Tao Zhang} },
	title        = { DeepSeekFakeNews-LLM-7B-Chat (Revision ccc44ec) },
	year         = 2025,
	url          = { https://huggingface.co/zt-ai/DeepSeekFakeNews-LLM-7B-Chat },
	doi          = { 10.57967/hf/4631 },
	publisher    = { Hugging Face }
}

7. License

This code repository is licensed under the MIT License. The use of DeepSeekFakeNews-LLM models is subject to the Model License. DeepSeekFakeNews-LLM supports commercial use.

<!-- See the [LICENSE-MODEL](https://github.com/deepseek-ai/deepseek-LLM/blob/main/LICENSE-MODEL) for more details. -->

8. Contact

If you have any questions, please raise an issue or contact us at [email protected] or [email protected].