---
title: HssaLLM-8B
canonical_url: "https://www.modelscope.cn/models/njauzwh/HssaLLM-8B"
md_url: "https://www.modelscope.cn/models/njauzwh/HssaLLM-8B.md"
repository: njauzwh/HssaLLM-8B
last_updated: 2026-05-24
license: "Apache License 2.0"
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - qwen3
architectures:
  - Qwen3ForCausalLM
parameters: 8.2B
tensor_type:
  - BF16
library_name:
  - safetensors
  - pytorch
frameworks:
  - Pytorch
downloads: 917
stars: 7
---

# HssaLLM-8B

> HssaLLM-8B - njauzwh 在 ModelScope 开源的模型。Large Language Model for Humanities and Social Sciences Academia

njauzwh/HssaLLM-8B 是 ModelScope 魔搭社区上的 8.2B 参数text-generation模型，采用 Apache License 2.0 许可。

- **Repository**: njauzwh/HssaLLM-8B
- **License**: Apache License 2.0
- **Tasks**: text-generation
- **Parameters**: 8.2B
- **Downloads**: 917
- **Stars**: 7
- **Last updated**: 2026-05-24

Source: https://www.modelscope.cn/models/njauzwh/HssaLLM-8B

---

<details open name="language">
<summary><strong>中文</strong></summary>

# 兰章大语言模型

<div align="center">
  <img src="./images/logo_new.png" alt="兰章大语言模型 Logo" width="450" />
</div>

# Large Language Model for Humanities and Social Sciences Academia

[ModelScope](https://www.modelscope.cn/models/njauzwh/HssaLLM-8B) | [License](https://www.apache.org/licenses/LICENSE-2.0)

**兰章大语言模型** 是首个开源的人文社会科学学术大语言模型，也是基于目前规模最大、学科覆盖最完整和时间跨度最长的人文社会科学学术语料所训练的开源模型。针对通用大模型在人文社科垂直领域缺乏深度语义理解、理论支撑不足以及长文本逻辑推演能力受限等问题，团队以 **Qwen3** 系列（8B/32B）为基座，构建了包含 **116亿 token** 的高质量双语学术语料数据集。通过**继续预训练**与**多任务指令微调**，兰章大语言模型深度内化了人文社会科学的学科知识体系与学术表达范式，旨在赋能新文科的“数智化”发展，为跨学科研究提供智能化的方法论指引与数据支持。

## 模型下载

兰章系列模型已在 ModelScope 社区开源，包含基座模型（Base）和指令微调模型（Instruct）。

| 模型名称 | 参数量 | 模型类型 | 下载地址 |
| --- | --- | --- | --- |
| **HssaLLM-8B-Base** | 8B | 基座模型 | [ModelScope Link](https://www.modelscope.cn/models/njauzwh/HssaLLM-8B-Base) |
| **HssaLLM-8B** | 8B | 指令模型 | [ModelScope Link](https://www.modelscope.cn/models/njauzwh/HssaLLM-8B) |
| **HssaLLM-32B-Base** | 32B | 基座模型 | [ModelScope Link](https://www.modelscope.cn/models/njauzwh/HssaLLM-32B-Base) |
| **HssaLLM-32B** | 32B | 指令模型 | [ModelScope Link](https://www.modelscope.cn/models/njauzwh/HssaLLM-32B) |

## 评价基准

为选取性能优越的基线模型并验证兰章大语言模型的有效性，本研究制定了面向人文社会科学学术领域的专用评价体系。针对学术研究场景，构建了如下评测任务：

| 任务 | 说明 | 测试数据量 |
| --- | --- | --- |
| 学术论文章节标题生成 | 基于学术全文本内容，生成论文章节标题 | 1000（中）+1000（英） |
| 学术论文结语生成 | 基于学术全文本内容，生成论文结论 | 1000（中）+1000（英） |
| 学术论文文献综述生成 | 基于学术全文本内容，生成论文文献综述 | 1000（中）+1000（英） |
| 学术文献实体识别 | 根据学术全文本内容进行实体标注 | 1000（中）+1000（英） |
| 图书自动分类 | 基于图书全文本数据判定学科类别 | 1000 |

## 数据收集与模型训练

### 数据来源与规模

本研究构建了人文社会科学学术数据集，总规模约 **116亿 token**，时间跨度覆盖 **1992-2025年**。

* **中文学术数据**：来源于中文社会科学引文索引（CSSCI）、CNKI人文社科期刊、人大复印报刊资料等，包含题录及学术全文本。
* **英文学术数据**：来源于 Web of Science (SSCI, A&HCI)、Project MUSE 和 Project Gutenberg，涵盖期刊摘要、论文全文本及经典学术专著。

### 训练策略

为解决通用模型在特定领域的“灾难性遗忘”并提升指令跟随能力，研究采用了两阶段训练策略：

1. **继续预训练 (Continued Pre-training)**：利用构建的 116B token 领域语料，对 **Qwen3-8B** 和 **Qwen3-32B** 进行全参数训练，注入人文社科特有的理论体系、专业术语及长文本逻辑结构。
2. **指令微调 (Instruction Tuning)**：构建高质量的多任务指令微调数据集（[hss_sft.json](https://www.modelscope.cn/datasets/njauzwh/HssaLLM-SFT-data)），涵盖学术论文章节标题生成、学术论文结语生成、学术论文文献综述生成、学术文献实体识别、图书自动分类、学科分类、语步识别、术语抽取与分类、中英文献互译等任务。

## 快速使用

```python
from modelscope import AutoModelForCausalLM, AutoTokenizer

model_name = "njauzwh/HssaLLM-8B"

# 加载模型与分词器
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype="auto",
    trust_remote_code=True,
)

# 构造提示词
prompt = "简述大语言模型在数字人文中的应用前景。"
messages = [{"role": "user", "content": prompt}]

# 推理生成
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(model_inputs.input_ids, max_new_tokens=512)
generated_ids = [
    output_ids[len(input_ids):]
    for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)
```

## 研究团队

兰章大语言模型由南京理工大学沈思教授团队研发。如果大家有什么问题，请通过如下邮箱进行交流：jisuanyuyan@163.com。谢谢大家。

## 声明

1. **数据合规性**：从纯粹学术研究的目的，兰章大语言模型基于大规模人文社会科学学术数据训练，我们已采取技术手段对数据进行清洗与脱敏。
2. **幻觉问题**：受大语言模型架构限制，模型生成的学术内容（如引文、数据等）可能存在“幻觉”现象。在用于严谨学术研究时，**务必对生成内容进行人工核实**。
3. **免责条款**：本模型仅供学术研究与交流使用。任何机构或个人因使用本模型及其衍生内容而引发的直接或间接损失，研发团队不承担法律责任。使用本模型需严格遵守《生成式人工智能服务管理暂行办法》及相关法律法规。

</details>

<details name="language">
<summary><strong>English</strong></summary>

# Lanzhang Large Language Model

<div align="center">
  <img src="./images/logo_new.png" alt="Lanzhang Large Language Model Logo" width="450" />
</div>

# Large Language Model for Humanities and Social Sciences Academia

[ModelScope](https://www.modelscope.cn/models/njauzwh/HssaLLM-8B) | [License](https://www.apache.org/licenses/LICENSE-2.0)

**Lanzhang Large Language Model** is the first open-source academic large language model for the humanities and social sciences. It is also an open-source model trained on one of the largest, most discipline-complete, and longest-spanning academic corpora in the humanities and social sciences to date.

To address the limitations of general-purpose large language models in vertical humanities and social sciences scenarios, including insufficient deep semantic understanding, limited theoretical grounding, and constrained long-context reasoning, our team built a high-quality bilingual academic corpus containing **11.6 billion tokens** based on the **Qwen3** series models (8B/32B). Through **continued pre-training** and **multi-task instruction tuning**, Lanzhang deeply internalizes the disciplinary knowledge systems and academic expression patterns of the humanities and social sciences. The model is designed to support the digital and intelligent development of the new liberal arts and provide intelligent methodological guidance and data support for interdisciplinary research.

## Model Downloads

The Lanzhang model series has been released on the ModelScope community, including both base models and instruction-tuned models.

| Model Name | Parameters | Model Type | Download |
| --- | --- | --- | --- |
| **HssaLLM-8B-Base** | 8B | Base Model | [ModelScope Link](https://www.modelscope.cn/models/njauzwh/HssaLLM-8B-Base) |
| **HssaLLM-8B** | 8B | Instruction Model | [ModelScope Link](https://www.modelscope.cn/models/njauzwh/HssaLLM-8B) |
| **HssaLLM-32B-Base** | 32B | Base Model | [ModelScope Link](https://www.modelscope.cn/models/njauzwh/HssaLLM-32B-Base) |
| **HssaLLM-32B** | 32B | Instruction Model | [ModelScope Link](https://www.modelscope.cn/models/njauzwh/HssaLLM-32B) |

## Evaluation Benchmark

To select strong baseline models and verify the effectiveness of Lanzhang, this study designed a dedicated evaluation framework for academic scenarios in the humanities and social sciences. The benchmark includes the following tasks:

| Task | Description | Test Set Size |
| --- | --- | --- |
| Academic paper section title generation | Generate section titles based on the full text of academic papers | 1,000 Chinese + 1,000 English |
| Academic paper conclusion generation | Generate conclusions based on the full text of academic papers | 1,000 Chinese + 1,000 English |
| Academic paper literature review generation | Generate literature reviews based on the full text of academic papers | 1,000 Chinese + 1,000 English |
| Academic literature entity recognition | Perform entity annotation based on the full text of academic literature | 1,000 Chinese + 1,000 English |
| Automatic book classification | Determine subject categories based on full-text book data | 1,000 |

## Data Collection and Model Training

### Data Sources and Scale

This study constructed an academic dataset for the humanities and social sciences with a total scale of approximately **11.6 billion tokens**, covering the period from **1992 to 2025**.

* **Chinese academic data**: Collected from the Chinese Social Sciences Citation Index (CSSCI), CNKI humanities and social sciences journals, and the Information Center for Social Sciences of Renmin University of China, including bibliographic records and full-text academic materials.
* **English academic data**: Collected from Web of Science (SSCI, A&HCI), Project MUSE, and Project Gutenberg, covering journal abstracts, full-text papers, and classic academic monographs.

### Training Strategy

To mitigate catastrophic forgetting in general-purpose models and improve instruction-following capability in specialized academic domains, this research adopts a two-stage training strategy:

1. **Continued Pre-training**: Using the constructed 11.6B-token domain corpus, we perform full-parameter training on **Qwen3-8B** and **Qwen3-32B**, injecting theoretical systems, domain terminology, and long-text logical structures specific to the humanities and social sciences.
2. **Instruction Tuning**: We construct a high-quality multi-task instruction-tuning dataset ([hss_sft.json](https://www.modelscope.cn/datasets/njauzwh/HssaLLM-SFT-data)), covering academic paper section title generation, academic paper conclusion generation, academic paper literature review generation, academic literature entity recognition, book classification, discipline classification, move recognition, terminology extraction and classification, and Chinese-English academic translation.

## Quick Start

```python
from modelscope import AutoModelForCausalLM, AutoTokenizer

model_name = "njauzwh/HssaLLM-8B"

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype="auto",
    trust_remote_code=True,
)

# Build prompt
prompt = "Briefly describe the application prospects of large language models in digital humanities."
messages = [{"role": "user", "content": prompt}]

# Generate response
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(model_inputs.input_ids, max_new_tokens=512)
generated_ids = [
    output_ids[len(input_ids):]
    for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)
```

## Research Team

Lanzhang Large Language Model was developed by Professor Shen Si's team at Nanjing University of Science and Technology. If you have any questions, please contact us by email: jisuanyuyan@163.com. Thank you.

## Statement

1. **Data Compliance**: For purely academic research purposes, Lanzhang Large Language Model is trained on large-scale academic data in the humanities and social sciences. We have applied technical measures for data cleaning and desensitization.
2. **Hallucination Risk**: Due to the inherent limitations of large language model architectures, the academic content generated by the model, such as citations and data, may contain hallucinations. For rigorous academic research, generated content **must be manually verified**.
3. **Disclaimer**: This model is intended only for academic research and communication. The research team assumes no legal responsibility for any direct or indirect losses caused by the use of this model or its derivative content. Users must strictly comply with the Interim Measures for the Management of Generative Artificial Intelligence Services and other relevant laws and regulations.

</details>