by Charlotte Johnson | Jul 18, 2024 | AI Testing, Blog, Featured |
Large Language Model (LLM) software testing requires a different approach compared to conventional mobile, web, and API testing. This is due to the fact that the output of such LLM or AI applications is unpredictable. A simple example is that even if you give the same prompt twice, you will receive unique outputs from the LLM model. We faced similar challenges when we ventured into GenAI development. So based on our experience of testing the AI applications we developed and other LLM testing projects we have worked on, we were able to develop a strategy for testing AI and LLM solutions. So in this blog, we will be helping you get a comprehensive understanding of LLM software testing.
LLM Software Testing Approach
By identifying the quality problems associated with LLMs, you can effectively strategize your LLM software testing approach. So let’s start by comprehending the prevalent LLM quality and safety concerns and learn how to find them with LLM quality checks.
Hallucination
As the word suggests, Hallucination is when your LLM application starts providing irrelevant or nonsensical responses. It is in reference to how humans hallucinate and see things that do not exist in real life and think them to be real.
Example:
Prompt: How many people are living on the planet Mars?
Response: 50 million people are living on Mars.
How to Detect Hallucinations?
Given that the LLM can hallucinate in multiple ways for different prompts, detecting these hallucinations is a huge challenge that we have to overcome during LLM software testing. We recommend using the following methods,
Check Prompt-Response Relevance – Checking the relevance between a given prompt and response can assist in recognizing hallucinations. We can use the BLEU scoreBLEU scoreMeasures how closely a generated text matches reference texts by comparing short sequences of words and BERT scoreBERT scoreAssesses how similar a generated text is to reference texts by comparing their meanings using BERT language model embeddings to check the relevance between prompt and LLM response.
- BLEU score is calculated with exact matching by utilizing the Python Evaluate library. The score ranges from 0 to 1 and a higher score indicates a greater similarity between your prompt and response.
- BERT score is calculated with semantic matching and it is a powerful evaluation metric to measure text similarity.
Check Against Multiple Responses – We can check the accuracy of the actual response by comparing it to various randomly generated responses for a given prompt. We can use Sentence Embedding Cosine Distance & LLM Self-evaluation to check the similarity.
Testing Approach
- Shift Left Testing – Before deploying your LLM application, evaluate your model or RAG implementation thoroughly
- Shift Right Testing – Check BERT score for production prompts and responses
Prompt Injections
Jailbreak – Jailbreak is a direct prompt injection method to get your LLM to ignore the established safeguards that tell the system what not to do. Let’s say a malicious user asks a restricted question in the Base64 formatBase64 formatIt is a way of encoding binary data into a text format using a set of 64 different ASCII characters , your LLM application should not answer the question. Security experts have already identified various Jailbreaking methods in commonly used LLMs. So it is important to analyze such methods and ensure your LLM system is not affected by them.
Indirect Injection
- Hidden prompts are often added by attackers in your original prompt.
- Attackers intentionally make the model to get data from unreliable sources. Once training data is incorrect, the response from LLM will also be incorrect.
Refusals – Let’s say your LLM model refuses to answer for a valid prompt, it could be because the prompt might be modified before sending it to LLM.
How to prevent Prompt Injection?
- Ensure your training data doesn’t have sensitive information
- Ensure your model doesn’t get data from unreliable external sources
- Perform all the security checks for LLM APIs
- Check substrings like (Sorry, I can’t, I am not allowed) in response to detect refusals
- Check response sentiment to detect refusals
RAG Injection
RAG is an AI framework that can effectively retrieve and incorporate outside information with the prompt provided to LLM. This allows the model to generate an accurate response when contextual cues are given by the user. The outside or external information is usually retrieved and stored in a vector database.
If poisoned data is obtained from an external source, how will LLM respond? Clearly, your model will start producing hallucinated responses. This phenomenon in LLM software testing is referred to as RAG injection.
Data Leakage
Data Leakage occurs when confidential or personal information is exposed either through a Prompt or LLM response.
Data Leak from Prompt – Let’s assume a user mentions their credit card number or password in their prompt. In that case, the LLM application must identify this information to be confidential even before it sends the request to the model for processing.
Data Leak from Response – Let’s take a Healthcare LLM application as an example here. Even if a user asks for medical records, the model should never disclose sensitive patient information or personal data. The same applies to other types of LLM applications as well.
How to prevent Data Leakage?
- Ensure training data doesn’t store any personal or confidential information.
- Use Regex to check all the incoming prompts or outgoing responses for Personal Identifiable Information.(PII)
Grounding Issues
Grounding is a method for tailoring your LLM to a particular domain, persona, or use case. We can cover this in our LLM software testing approach through prompt instructions. When an LLM is limited to a specific domain, all of its responses must fall within that domain. So manual testers have a vital responsibility here in identifying any LLM grounding problems.
Testing Approach
- Ask multiple questions that are not relevant to the Grounding instructions.
- Add an active response monitoring mechanism in Production to check the Groundedness score.
Token Usage
There are numerous LLM APIs in the market that charge a fee for the tokens generated from the prompts. Let’s say your LLM application is generating more tokens after a new deployment, this will result in a surge in the monthly billing for API usage.
The pricing of LLM products for many companies is typically determined by Token consumption and other resources utilized. If you don’t calculate & monitor token usage, your LLM product will not make the expected revenue from it.
Testing Approach
- Monitor token usage and the monthly cost constantly.
- Ensure the response limit is working as expected before each deployment.
- Always look for optimizing token usage.
General LLM Software Testing Tips
For effective LLM software testing, there are several key steps that should be followed. The first step is to clearly define the objectives and requirements of your application. This will provide a clear roadmap for testing and help determine what aspects need to be focused on during the testing process
Moreover, continuous integration (CI) plays an important role in ensuring a smooth development workflow by constantly integrating new code into the existing codebase while running automated tests simultaneously. This helps catch any issues early on before they pile up into bigger problems.
It is crucial to have a dedicated team responsible for monitoring and managing quality assurance throughout the entire development cycle. A competent team will ensure effective communication between developers and testers resulting in timely identification and resolution of any issues found during testing.
Conclusion:
LLM software testing may seem like a daunting and time-consuming process, but it is an essential step in delivering a high-quality product to end-users. By following the steps outlined above, you can ensure that your LLM application is thoroughly tested and ready for success in the market. As it is an evolving technology, there will be rapid advancements in the way we approach LLM application testing. So make sure to keep updating your approach by keeping yourself updated. Also, make sure to keep an eye out on this space for more informative content.
by admin | Aug 16, 2021 | AI Testing, Blog, Latest Post |
In recent years organizations have invested significantly in structuring their testing process to ensure continuous releases of high-quality software. But all that streamlining doesn’t apply when artificial intelligence enters the equation. Since the testing process itself is more challenging, organizations are now in a dire need of a different approach to keep up with the rapidly increasing inclusion of AI in the systems that are being created. AI technologies are primarily used to enhance our experience with the systems by improving efficiency and providing solutions for problems that require human intelligence to solve. Despite the high complexity of the AI systems that increase the possibility of errors, we have been able to successfully implement our AI testing strategies to deliver the best software testing services to our clients. So in this AI Testing Tutorial, we’ll be exploring the various ways we can handle AI Testing effectively.
Understanding AI
Let’s start this AI Testing Tutorial with a few basics before heading over to the strategies. The fundamental thing to know about machine learning and AI is that you need data, a lot of data. Since data plays a major role in the testing strategy, you would have to divide it into three parts, namely test set, development set, and training set. The next step is to understand how the three data sets work together to train a neural network before testing your AI-based application.
Deep learning systems are developed by feeding several data into a neural network. The data is fed into the neural network in the form of a well-defined input and expected output. After feeding data into the neural network, you wait for the network to give you a set of mathematical formulae that can be used to calculate the expected output for most of the data points that you feed the neural network.
For example, if you were creating an AI-based application to detect deformed cells in the human body. The computer-readable images that are fed into the system make up the input data, while the defined output for each image forms the expected result. That makes up your training set.
Difference between Traditional systems and AI systems
It is always smart to understand any new technology by comparing it with the previous technology. So we can use our experience in testing the traditional systems to easily understand the AI systems. The key to that lies in understanding how AI systems differ from traditional systems. Once we have understood that, we can make small tweaks and adjustments to the already acquired knowledge and start testing AI systems optimally.
Traditional Software Systems
Features:
Traditional software is deterministic, i.e., it is pre-programmed to provide a specific output based on a given set of inputs.
Accuracy:
The accuracy of the software depends upon the developer’s skill and is deemed successful only if it produces an output in accordance with its design.
Programming:
All software functions are designed based on loops and if-then concepts to convert the input data to output data.
Errors:
When any software encounters an error, remediation depends on human intelligence or a coded exit function.
AI Systems:
Now, we will see the contrast of the AI systems over the traditional system clearly to structure the testing process with the knowledge gathered from this understanding.
Features:
Artificial Intelligence/machine learning is non – deterministic, i.e., the algorithm can behave differently for different runs since the algorithms are continuously learning.
Accuracy:
The accuracy of AI learning algorithms depends on the training set and data inputs.
Programming:
Different input and output combinations are fed to the machine based on which it learns and defines the function.
Errors:
AI systems have self-healing capabilities whereby they resume operations after handling exceptions/errors.
From the difference between each topic under the two systems we now have a certain understanding with which we can make modifications when it comes to testing an AI-based application. Now let’s focus on the various testing strategies in the next phase of this AI Testing Tutorial.
Testing Strategy for AI Systems
It is better not to use a generic approach for all use cases, and that is why we have decided to give specific test strategies for specific functionalities. So it doesn’t matter if you are testing standalone cognitive features, AI platforms, AI-powered solutions, or even testing machine learning-based analytical models. We’ve got it all covered for you in this AI Testing Tutorial.
Testing standalone cognitive features
Natural Language Processing:
1. Test for ‘precision’ – Return of the keyboard, i.e., a fraction of relevant instances among the total retrieved instances of NLP.
2. Test for ‘recall’ – A fraction of retrieved instances over the total number of retrieved instances available.
3. Test for true positives, True negatives, False positives, False negatives. Confirm that FPs and FNs are within the defined error/fallout range.
Speech recognition inputs:
1. Conduct basic testing of the speech recognition software to see whether the system recognizes speech inputs.
2. Test for pattern recognition to determine if the system can identify when a unique phrase is repeated several times in a known accent and whether it can identify the same phrase when repeated in a different accent.
3. Test how speech translates to the response. For example, a query of “Find me a place where I can drink coffee” should not generate a response with coffee shops and driving directions. Instead, it should point to a public place or park where one can enjoy coffee.
Optical character recognition:
1. Test the OCR and Optical word recognition basics by using character or word input for the system to recognize.
2. Test supervised learning to see if the system can recognize characters or words from printed, written or cursive scripts.
3. Test deep learning, i.e., check whether the system can recognize the characters or words from skewed, speckled, or binarized documents.
4. Test constrained outputs by introducing a new word in a document that already has a defined lexicon with permitted words.
Image recognition:
1. Test the image recognition algorithm through basic forms and features.
2. Test supervised learning by distorting or blurring the image to determine the extent of recognition by the algorithm.
3. Test pattern recognition by replacing cartoons with the real image like showing a real dog instead of a cartoon dog.
4. Test deep learning scenarios to see if the system can find a portion of an object in a large image canvas and complete a specific action.
Now we will be focusing on the various strategies for algorithm testing, API integration, and so on in this AI Testing Tutorial as they are very important when it comes to testing AI platforms.
Algorithm testing:
1. Check the cumulative accuracy of hits (True positives and True negatives) over misses (False positives and False negatives)
2. Split the input data for learning and algorithm.
3. If the algorithm uses ambiguous datasets in which the output for a single input is not known, then the software should be tested by feeding a set of inputs and checking if the output is related. Such relationships must be soundly established to ensure that the algorithm doesn’t have defects.
4. If you are working with an AI which involves neural networks, you have to check it to see how good it is with the mathematical formulae that you have trained it with and how much it has learned from the training. Your training algorithm will show how good the neural network algorithm is with its result on the training data that you fed it with.
The Development set
However, the training set alone is not enough to evaluate the algorithm. In most cases, the neural network will correctly determine deformed cells in images that it has seen several times. But it may perform differently when fed with fresh images. The algorithm for determining deformed cells will only get one chance to assess every image in real-life usage, and that will determine its level of accuracy and reliability. So the major challenge is knowing how well the algorithm will work when presented with a new set of data that it isn’t trained on.
This new set of data is called the development set. It is the data set that determines how you modify and adjust your neural network model. You adjust the neural network based on how well the network performs on both the training and development sets, this means that it is good enough for day-to-day usage.
But if the data set doesn’t do well with the development set, you need to tweak the neural network model and train it again using the training set. After that, you need to evaluate the new performance of the network using the development set. You could also have several neural networks and select one for your application based on its performance on your development set.
API integration:
1. Verify the input request and response from each application programming interface (API).
2. Conduct integration testing of API and algorithms to verify the reconciliation of the output.
3. Test the communication between components to verify the input, the response returned, and the response format & correctness as well.
4. Verify request-response pairs.
Data source and conditioning testing:
1. Verify the quality of data from the various systems by checking their data correctness, completeness & appropriateness along with format checks, data lineage checks, and pattern analysis.
2. Test for both positive and negative scenarios.
3. Verify the transformation rules and logic applied to the raw data to get the output in the desired format. The testing methodology/automation framework should function irrespective of the nature of the data, be it tables, flat files, or big data.
4. Verify if the output queries or programs provide the intended data output.
System regression testing:
1. Conduct user interface and regression testing of the systems.
2. Check for system security, i.e., static and dynamic security testing.
3. Conduct end-to-end implementation testing for specific use cases like providing an input, verifying data ingestion & quality, testing the algorithms, verifying communication through the API layer, and reconciling the final output on the data visualization platform with the expected output.
Testing of AI-powered solutions
In this part of the AI Testing Tutorial, we will be focusing on strategies to use when testing AI-powered solutions.
RPA testing framework:
1. Use open-source automation or functional testing tools such as Selenium, Sikuli, Robot Class, AutoIT, and so on for multiple purposes.
2. Use a combination of pattern, text, voice, image, and optical character recognition testing techniques with functional automation for true end-to-end testing of applications.
3. Use flexible test scripts with the ability to switch between machine language programming (which is required as an input to the robot) and high-level language for functional automation.
Chatbot testing framework:
1. Maintain the configurations of basic and advanced semantically equivalent sentences with formal & informal tones, and complex words.
2. Generate automated scripts in python for execution.
3. Test the chatbot framework using semantically equivalent sentences and create an automated library for this purpose.
4. Automate an end-to-end scenario that involves requesting for the chatbot, getting a response, and finally validating the response action with accepted output.
Testing ML-based analytical models
Analytical models are built by the organization for the following three main purposes.
Descriptive Analytics:
Historical data analysis and visualization.
Predictive Analytics:
Predicting the future based on past data.
Prescriptive Analytics:
Prescribing course of action from past data.
Three steps of validation strategies are used while testing the analytical model:
1. Split the historical data into test & train datasets.
2. Train and test the model based on generated datasets.
3. Report the accuracy of the model for the various generated scenarios as well.
All types of testing are similar:
It’s natural to feel overwhelmed after seeing such complexity. But as a tester, if one is able to see through the complexity, they will be able to that the foundation of testing is quite similar for both AI-based and traditional systems. So what we mean by this is that the specifics might be different, but the processes are almost identical.
First, you need to determine and set your requirements. Then you need to assess the risk of failure for each test case before running tests and determining if the weighted aggregated results are at a predefined level or above the predefined level. After that, you need to run some exploratory testing to find biased results or bugs as in regular apps. Like we said earlier, you can master AI testing by building on your existing knowledge.
With all that said, we know for a fact that an AI-based system provides a highly functional dynamic output with the same input when it is run again and again since the ML algorithm is a learning algorithm. Also, most of the applications today have some type of Machine Learning functionality to enhance the relationship of the applications with the users. AI inclusion on a much larger scale is inevitable as we humans will stop at nothing until the software we create has human-like functionalities. So it’s necessary for us to adapt to the progress of this AI revolution.
Conclusion:
We hope that this AI Testing Tutorial has helped you understand the AI algorithms and their nature that will enable you to tailor your own test strategies and test cases that cater to your needs. Applying out-of-the-box thinking is crucial for testing AI-based applications. As a leading QA company, we always implement the state of the art strategies and technologies to ensure quality irrespective of the software being AI-based or not.
by admin | May 16, 2019 | AI Testing, Blog |
The use of AI (Artificial Intelligence) in software testing is one of the latest emerging trends in the software industry. The main aim behind the application of AI to software testing tools is to make the software development lifecycle easier.
With testing being a crucial process in the software development lifecycle (SDLC), the use of Artificial Intelligence in software testing can significantly streamline the testing process, making it smarter and more efficient. The deployment of a smarter testing process is crucial for any top software company due to the transformation in DevOps and the frequent release of new software and products. Hence, it is widely projected that AI will play a vital role in software testing in the years ahead for a number of reasons, some of which are mentioned below:
A software testing company that uses AI as part of its software testing methodologies enables its testers to move beyond the traditional models of testing. The use of AI involves the assimilation of machines that are capable of meticulous replication of human behaviour. Hence, AI in software testing can ensure that the automated testing process is even more precise, robust and continuous.
The amount of tedious and mundane tasks (though important) in software testing can be reduced with the help of AI. In addition, AI also facilitates the automation of the testing process through the application of reasoning and problem solving. The ‘machine learning’ subset of AI is also used in some cases for applying algorithms that automatically enhance the testing tool via the collection of massive amounts of data generated through testing.
The amalgamation of AI in the creation and execution of software tests, as well as data analysis, can simplify the overall testing process for the testers. When AI is applied to software testing, testers no longer need to update test cases manually and repeatedly. Moreover, AI tools also give testers the capability to identify controls more effectively and observe the connection between defects and components.
AI in software testing requires data, computing power and algorithms. AI can enhance automation testing and is used widely for the purpose of object application categorization for a variety of user interfaces. Such a scenario paves the way for classification of recognized controls during tool creation, thereby enabling testers to pre-train certain controls that are a component of out-of-the-box setups.
An automated and continuous testing platform powered by AI has the capability to recognize changed controls more efficiently as compared to manual testers. As a result of constant updates to algorithms, it is possible for testers to monitor even the slightest changes. As a result software testing services are becoming increasingly inclined towards AI since software test automation tools enabled by AI have the ability to provide enhanced value to testers.
Most of the test automation tools have the ability to run only a few predetermined tests since they are unable to determine on their own which tests to run. However, the application of AI in software testing can improve the testing ability of the tools by enabling them to make decisions to run tests based on changing data. An AI-enabled bot can decide which tests to run, and subsequently run them post reviewing the current test status, code coverage, recent code changes, and other metrics.
AI in automated testing can lead to a significant increase in the overall scope and depth of testing and, thus, improve the quality of software. Automated testing can easily assess whether the software meets the expectations, by scanning the memory and file contents, the state of internal program and data tables. The use of AI can help automation testing provide better test coverage given its ability to execute more than 1000 different test cases in each test run.
By applying AI in software testing, a software company can achieve its perceived ‘quality’ goals. AI is apparently set to become a vital part of the quality engineering process of the future because it can be applied to diverse actions. These actions include prioritizing testing, enhancing automation, optimizing test cases, reducing mundane analysis tasks, and improving User Interface testing.
The applications of AI in software testing can give an extraordinary boost to the overall effectiveness of software testing tool suites. The ultimate objective behind the use of AI in software testing is to help testers test their code more efficiently, and create high-quality software at a faster pace. AI in automated testing can especially enable the testers to eliminate repetitive, time-consuming manual tests and allow them to create new, complex automated software tests with advanced features – connect with us today to speak with our experts.
