ChroKnowledge: Unveiling Chronological Knowledge of Language Models in Multiple Domains

Overview

⏰ ChroKnowledge is a research framework designed to evaluate and update the chronological knowledge of large language models (LLMs). It builds on the ChroKnowBench dataset, which enables testing LLMs' ability to handle chronologically accumulated knowledge across multiple domains, including general, biomedical, legal, commonsense, and mathematical facts. This repository also features ChroKnowPrompt, a technique for in-depth prompting to enhance temporal reasoning and improve the accuracy of LLMs over a timeline.

🏠 Homepage | 📃 Paper | 🤗 Datasets

Updates

[Nov 28, 2024] Our paper is now updated to the latest version.
[Oct 18, 2024] We post a homepage for overview of our work.
[Oct 11, 2024] We have released the code and data.

Task Definition

Our knowledge categorization is based on knowledge classification based on the Sampling-based Knowledge Categorization (SliCK) framework by Gekhman et al: Sampling the model's answer to question using various few-shot exemplar sets D. The sampling is conducted under two temperature conditions: $\tau$ = 0 and $\tau$ > 0. Then categorizes the degree to wich the model knows each piece of knowledge into four levels: HighlyKnown, MaybeKnown, WeaklyKnown, and Unknown. We modify it into three categories as defined below. Knowledge is categorized into:

• Correct: All objects generated are included in the answer set.
• Partial Correct: At least one generated object is in the answer set.
• Incorrect: None of the generated objects are included in the answer set.

Category	Definition	Description
Correct	$( { \hat{o_i} \mid M(D_i,s,r,t) = \hat{o_i}; M, \tau=0 }_{i=1}^{n} \subseteq A )$	All objects generated with greedy decoding are entirely included within the answer set.
Partial Correct	$( \bigcup_{\tau \in \mathcal{T}} { \hat{o_i} \mid M(D_i, s, r, t) = \hat{o_i}; M, \tau }_{i=1}^{n} \cap A \neq \emptyset )$	At least one generated object from greedy decoding or temperature sampling is in the answer set.
Incorrect	$( \bigcup_{\tau \in \mathcal{T}} { \hat{o_i} \mid M(D_i, s, r, t) = \hat{o_i}; M, \tau }_{i=1}^{n} \cap A = \emptyset )$	None of the generated objects, either from greedy decoding or temperature sampling, are included in the answer set.

Datasets

Our ChroKnowBench consists of datasets from multiple domains, with characteristics as shown in below:

• Time-variant Knowledge: General, biomedical, and legal datasets, containing facts that change over time. It is divided into two temporal state:
  • dynamic, where knowledge has evolved over the accumulated period.
  • static, where no change occurred during the accumulated period though it has potential to be changed.
• Time-invariant Knowledge: Commonsense and mathematics datasets, containing facts that remain constant.

Time Dependency	Domain (Time Frame)	# of Relations	Structured	Format	Temporal State	# of Examples	Source
Time Variant	General (2010-2023)	8	Yes	(s, r, o, t)	Dynamic	8,330	Wikidata
					Static	8,302	Wikidata
	Biomedical (2020-2024)	14	Yes	(s, r, o, t)	Dynamic	7,345	UMLS
					Static	7,345	UMLS
	Legal (2010-2023)	6*	No	QA	Dynamic	3,142	CFR
					Static	3,142	CFR
Time Invariant	Commonsense	8	Yes	(s, r, o)	Invariant	24,788	CSKG
	Math	12	Yes	(s, r, o)	Invariant	2,585	Math-KG

* For LEGAL dataset in time variant, it is the number of category like `Organization', as it is unstructured dataset without specific short relations.

Dataset Download

For Biomedical dataset(Dynamic, Static and Fewshot), it is released here. Please append each jsonl file to appropriate directory(/ChroKnowBench for Dynamic, Static file, and /ChroKnowBench/Fewshots for Fewshot file). Other benchmarks can be downloaded in Huggingface Datasets.

Installation

To ensure compatibiliity with other libraries, we recommend using the folliwng versions. You can adjust it based on your environments:

• Python 3.10.14
• PyTorch 2.4.0
• CUDA 12.2

Then, follow the order of installation.

1. Clone the repository:

   git clone https://github.com/dmis-lab/ChroKnowledge.git
   cd ChroKnowledge

2. Install dependencies:

   pip install -r requirements.txt

3. Download the dataset from HuggingFace:

   git clone https://huggingface.co/datasets/dmis-lab/ChroKnowBench

   Please download the dataset in Huggingface first, then download Biomedical datset from the link above.
   📌 We currently do not support load_datset module. Please wait for an update.

Implementation

Step 1: Chronological Knowledge Check

To evaluate the initial temporal knowledge of the model, you can run the run_knowledge_check.sh script. This script uses a variety of language models, including both open-source and proprietary LLMs, to assess their ability to recall time-sensitive information from different domains of ChroKnowBench. Below is an explanation of the key components:

1. Set CUDA Devices: The environment variable CUDA_VISIBLE_DEVICES allows you to specify which GPUs to use for running the script.

   export CUDA_VISIBLE_DEVICES=0

   Adjust this variable according to the GPU resources available.

2. Model Selection: The script includes multiple language models such as Llama3.1_8B and gpt-4o-mini. You can change the model_name variable to choose the model you wish to use. If you use OpenAI models, put your api key into 'api_key = "YOUR API KEY" in sources/utils.py file.

   model_name="Llama3.1_8B"

   ⚠️ Warning: As the basic sampling is five times for each year, be careful for the cost of calling API.

3. Domain and Template: You can specify the domain (e.g., General, Biomedical, Legal) and the type of task (generation, QA) by setting the domain and template variables, respectively.

   domain="General"
   template="generation"

4. Temperature and Decoding: The script runs the model with different temperature settings (0.0, 0.7) to capture variations in predictions. You can change the temperature variable for different behaviors. Please do both temperature settings for evaluation.

   temperature=0.0

5. Running the Script: After setting up, you can execute the following command to run the model evaluation:

   CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES python ChroKnowledge.py \
       --model_name $model_name \
       --dtype bfloat16 \
       --device_num auto \
       --gpu_util 0.90 \
       --multi_gpu $multi_gpu \
       --max_tokens 50 \
       --domain $domain \
       --template $template \
       --temperature $temperature \
       --save_results True

   This script helps to evaluate how well the language models can recall time-sensitive knowledge across different domains.

Evaluation

Use the run_classification.ipynb notebook to classify and analyze the model's temporal knowledge.

Step 2: ChroKnowPrompt Implementation

After running Step 1, you can use the ChroKnowPrompt prompting strategy to enhance the model's ability to recall and reason about chronological knowledge. This is implemented using the run_chroknow_prompt.sh script. Below is a detailed breakdown of the process:

1. Set CUDA Devices: Specify the GPU to use by setting CUDA_VISIBLE_DEVICES.

   export CUDA_VISIBLE_DEVICES=0

2. Model Selection: Choose the language model for this task by modifying the model_name variable.

   model_name="Llama3.1_8B"

3. Year Span for Temporal Reasoning: You can configure the prev_span and next_span variables to determine how far back or forward in time the model should reason during the evaluation.

   prev_span=3
   next_span=3

   In our setting, we used three year span for both previous and next years.

4. Running the Script: To run the script and evaluate the prompting strategy, use the following command:

   CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES python ChroKnowPrompt.py \
       --model_name $model_name \
       --dtype bfloat16 \
       --device_num auto \
       --gpu_util 0.90 \
       --multi_gpu $multi_gpu \
       --max_tokens 50 \
       --domain General \
       --temperature 0.0 \
       --prev_span $prev_span \
       --next_span $next_span \
       --save_results True

   This script helps the model better navigate chronological changes in the knowledge base, allowing for a more accurate understanding of time-sensitive information.

Evaluation

Use the run_evaluation.ipynb notebook to evaluate the effectiveness of ChroKnowPrompt in improving temporal knowledge recall.

Citation

@article{park2024ChroKnowledge,
  title={ChroKnowledge: Unveiling Chronological Knowledge of Language Models in Multiple Domains},
  author={Yein Park and Chanwoong Yoon and Jungwoo Park and Donghyeon Lee and Minbyul Jeong and Jaewoo Kang},
  journal={arXiv preprint arXiv:2410.09870},
  year={2024}
}

Contact

For any questions or issues, feel free to reach out to [522yein (at) korea.ac.kr].