HK-O1aw is a legal assistant designed to handle complex legal reasoning, specifically for the Hong Kong legal system. It is built using the Align-Anything framework and trained on the O1aw-Dataset., based on the LLaMA-3.1-8B model. The primary goal of HK-O1aw is to improve the reasoning and problem-solving abilities of large language models in the legal domain. Importantly, all training data, code, and prompts used for synthetic data generation have been open-sourced, facilitating research and collaboration within the community.
This model addresses the need for intelligent legal assistance in Hong Kong, where legal issues require in-depth analysis and precise reasoning. HK-O1aw integrates advanced O1-style reasoning capabilities, allowing it to perform complex legal analysis, understand context, identify precedents, and interpret statutes. As the first complex reasoning model tailored for Hong Kong‘s common law system, it is particularly valuable for improving legal services and education.
Currently, HK-O1aw assists users by providing detailed reasoning paths to legal problems and serves as an interactive tool for legal education, helping students understand complex concepts and practice legal drafting. Looking ahead, we plan continuous updates to enhance its deployment in professional settings, aiming to support lawyers with legal research, case analysis, and drafting legal opinions, thereby improving efficiency and accuracy in Hong Kong's legal framework.
Our model can be accessed here.
| Model Cases |
|---|
Case 1Prompt: Output: 理解紧急情况的例外 考虑其他可能性 通过分析，我们确认在非紧急情况下，医生通常需要病人或其家属的同意才能进行治疗。只有在紧急情况下，为了拯救生命而无需获得同意。 |
Case 2Prompt Output: 方法一：法律纠纷的可能性 方法二：物业管理和维修问题 方法三：政府收回土地的风险 通过不同的角度分析，我们可以确认没有书面协议存在多种潜在风险。这些包括法律纠纷、物业管理问题以及政府收回土地的风险。每种方法都从不同方面验证了这些风险，并且都指向相同的结论，即缺乏书面记录会导致严重后果。 |
Case 3Prompt: Output: 评估滑水作为高危运动 检查保险条款中的例外情况 考虑医疗和意外保险索偿影响 通过多种方法分析，我们可以得出结论：如果受伤事件符合意外的定义，并且不在保险合同中的例外情况内，则可能不会影响索偿。但如果被视为自招或疏忽，则可能会影响索偿。 |
Case 4Prompt: Output: 方法一：从法律程序角度分析 方法二：从沟通效率角度分析 方法三：从策略性角度分析 通过多种方法验证后，最终确认优秀状书需具备清晰、准确、全面、易读及策略性的特征。这些特征共同确保法庭能够有效理解案件背景并作出公正判决。 |
Case 5Prompt: Output: 探讨仲裁的优势 比较调解和仲裁 验证和排除其他可能性 通过重新审视各个步骤，确认调解和仲裁在国际贸易中的独特作用。它们提供了灵活性、保密性和法律保障，使得商业能够有效地管理风险并维护关系。 |
O1aw-Dataset is a comprehensive legal question-thought-answer dataset, designed to evaluate and enhance legal reasoning capabilities in language models. The dataset follows the O1-style format, featuring complex legal scenarios that require multi-step reasoning.
How was our dataset constructed? First, we crawled and cleaned raw legal materials from the internet, including Hong Kong e-Legislation. Then, we used GPT-4o to generate corresponding questions and thought-answer pairs based on the raw legal materials.
The dataset contains 15,959 question-thought-answer triples, each equipped with complete chain-of-thought annotations. All content is presented in Simplified Chinese and stored in a structured JSON format. The difficulty level of the questions in this dataset is intermediate to advanced for legal professionals and law school students.
The question types cover case analysis, legal application, explanation of legal concepts and so on. Each QTA triple includes detailed question prompt, a 3-5 step chain-of-thought reasoning process, and answer. The reasoning process involves multi-stage validation, reflective verification steps, and cross-case consistency checks, ensuring diversity in reasoning.
Our training dataset can be accessed here.
Here is our prompt template for Question generation:
SYSTEM_PROMPT: str = """
# Task
基于以下参考法律材料，生成至少{n}个法律问题。问题需要满足以下要求：
1. 复杂度要求：
- 问题的答案需要通过对参考法律材料的深入理解和分析才能得出
- 应该是一个开放性问题，需要进行推理和解释
- 不能是简单的是非题或事实性问题
2. 问题形式：
- 可以是案例分析题
- 可以是法律适用题
- 可以是法律概念解释题
3. 问题结构：
- 明确的问题陈述
4. 难度级别：
中等难度，适合法律专业学生或从业者思考和讨论
5. 输出格式：
请严格使用JSON格式输出，结构如下：
{{
"questions": [
{{
"id": 1,
"type": "案例分析/法律适用/概念解释...",
"question": "具体问题",
}},
{{
"id": 2,
...
}},
{{
"id": 3,
...
}}
]
}}
"""
USER_PROMPT: str = """
# 参考法律材料
{prompt}
# 提示
生成的问题应该与提供的参考法律材料直接相关，但是必须假装参考法律材料对你不可见。
请以JSON格式输出：
"""Here is our prompt template for Thought(COT) and Answer generation:
SYSTEM_PROMPT: str = """你是一个专家级的AI助手，能够逐步解释推理过程。你将收到一个问题和相关参考资料。你的任务是重构并展示通向正确答案的完整推理路径。
对于每个推理步骤，提供一个标题，描述你在该步骤中所做的事情，以及内容。但必须展示至少三种不同的方法或途径来得出该答案。
要求：
1. 使用3-5个推理步骤
2. 探索多种方法以达到答案
3. 通过不同的方法验证给定答案
4. 考虑潜在的替代答案并解释为何被拒绝
5. 你必须假装没有参考资料，只可以把参考资料当作自己的知识
6. 考虑你可能是错的，如果你的推理是错的，它会在哪里
7. 充分测试所有其他可能性。你可能会错
8. 当你说你正在重新检查时，请真正重新检查，并使用另一种方法进行，不要只是说你正在重新检查
以JSON格式回应，包含以下键：
- 'title': 当前推理步骤的描述
- 'content': 该步骤的详细解释
- 'next_action': 'continue' 或 'final_answer'
有效的JSON响应示例：
[
{{
"title": "分析给定信息",
"content": "首先，让我们检查问题，以识别将指导我们解决过程的关键要素……",
"next_action": "continue"
}},
{{
"title": "...",
"content": "...",
"next_action": "continue"
}},
...
{{
"title": "...",
"content": "...",
"next_action": "final_answer"
}}
]
"""
USER_PROMPT: str = """
# 问题：
{prompt}
# 参考资料：
{references}
请以JSON格式输出：
"""We use Align-Anything framework to conduct SFT training on Llama-3.1-8B. The training dataset and hyperparameters used are detailed below.
Specifically, in SFT training, Q, T, and A in the training dataset are segmented by reserved tokens in the tokenizer of Llama-3.1-8B. For more details, see template:
system_prompt: str = ''
user_prompt: str = '<|reserved_special_token_0|>{input}<|reserved_special_token_1|>\n'
assistant_thinking: str = '<|reserved_special_token_2|>{thinking}<|reserved_special_token_3|>\n'
assistant_answer: str = '<|reserved_special_token_4|>{answer}<|reserved_special_token_5|>'
template = system_prompt + user_prompt + assistant_thinking + assistant_answer
train_cfgs
```
# The deepspeed configuration
ds_cfgs: ds_z3_config.json
# Number of training epochs
epochs: 3
# Seed for random number generator
seed: 42
# Batch size per device for training
per_device_train_batch_size: 4
# Batch size per device for evaluation
per_device_eval_batch_size: 4
# The number of gradient accumulation steps
gradient_accumulation_steps: 16
# Whether to use gradient checkpointing
gradient_checkpointing: True
# Initial learning rate
learning_rate: 2.e-5
# Type of learning rate scheduler
lr_scheduler_type: cosine
# Ratio of warmup steps for learning rate
lr_warmup_ratio: 0.03
# Weight decay coefficient
weight_decay: 0.0
# Hyper-parameters for adam optimizer
adam_betas: [0.9, 0.95]
# Hyper-parameters for adam epsilon
adam_epsilon: 1.e-8
# Enable bfloat 16 precision
bf16: True
# Enable float 16 precision
fp16: False
# The strategy of evaluation, choosing form [epoch, steps]
eval_strategy: epoch
# The evaluation interval in step-wise evaluation case
eval_interval: 10
# The max norm of gradient
max_grad_norm: 1.0
```
data_cfgs
```
# Datasets to use for training
train_datasets: HKAIR-Lab/HK-O1aw-SFT-16k
# The split of train datasets
train_split: train
```
model_cfgs
```
# Pretrained model name or path
model_name_or_path: meta-llama/Llama-3.1-8B
# Whether to trust remote code
trust_remote_code: True
# The max token length
model_max_length: 2048
```
For other hyperparameters, we use the default value in the configure file.
In the planning, coming soon.
HK-O1aw is developed by the HKAIR research team, part of HKGAI. We specialize in developing reasoning techniques for large language models (LLMs). If you have any questions about HK-O1aw, feel free to reach out to us on the GitHub issue page.
Please cite the repo if you use the data or code in this repo.
@misc{HK-O1aw,
author = {HKAIR Lab},
title = {HK-O1aw Models: Leveraging O1 Slow Thinking in the Development of Hong Kong Legal Large Language Models},
year = {2024},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/HKAIR-Lab/HK-O1aw}},
}HK-O1aw model and dataset are released under Apache License 2.0.