Do Trillions Of Parameters Help In LLM Effectiveness?
AI, LLM, Parameters, and Testing - Venkat Ramakrishnan. Image Source: Unsplash

Do Trillions Of Parameters Help In LLM Effectiveness?

"The more, the merrier" - A great saying to reflect on while organizing a party. Does the same apply for the number of parameters in a large language model (LLM) in increasing its effectiveness?

The Number Of Parameters Game

The more parameters there are, the more connections can be made between the parameters, the goal being more meaningful associations in the parameters that would lead to the right answer. When I see release of LLMs that market for accuracy based on the no. of parameters, I tend to be become skeptical. 'Our LLM has a trillion parameters!' an LLM release note would say, not mentioning further if that trillion would be really effective in increasing LLM's accuracy in answers.

These parameters operate on data. If the underlying data is inaccurate, irrespective of the number of parameters, the output would be wrong. It all depends on what we train the neural network with. If we train the LLM with say The Internet, chances are that the LLM answers would not only be incorrect, but inconsistent too (meaning if I ask the same question tomorrow, I will get a different answer!).

Whereas, if I train the LLM on a focused domain like say car manufacturing, where I have control over what data I feed to the LLM, there is a much higher possibility of getting the right answer, and here, increasing the number of parameters help because the associations between the parameters are developed well as we have more parameters.

But that is only to a certain extent! There is a threshold above which increasing the number of parameters do not lead to more accurate results irrespective of how correct your data is! More parameters thereafter will only lead to overfitting - your model will not work for future data that is not represented already in the training data!

Software Testing and Quality Angle

What does this mean to a Software Testing of a Quality person while they test an LLM that has curated data (not the entire Internet) ?

  1. Validate the data - that's the first and foremost for quality. Keep diverse set of data relevant to your domain/business that need to be tested with. Keep integrating these data to your training dataset till you get a comprehensive dataset. Data augmentation helps in generating variations of data. This dataset should represent a focused task.
  2. Start testing with a some limited number of parameters. If you don't find the necessary accuracy with the results, increase the number of parameters and see if it helps.
  3. Have a performance chart for the LLM effectiveness vis-a-vis the number of parameters. You will find the threshold wherein increasing the number of parameters further would not increase the effectiveness of the LLM. Keep in mind the cost involved in increasing the parameters and have that data too in the chart.
  4. Test with various datasets that accomplish the same task and see if your model works well for them, and repeat steps 1 to 3.

Conclusion

Blindly increasing the number of parameters do not lead to LLM effectiveness and quality. We need to be conscious of data quality, the task that we are focused on, performance criteria, and cost effectiveness. An optimal balance is achieved by thoroughly testing the accuracy against the number of parameters.

Godwin Josh

Co-Founder of Altrosyn and DIrector at CDTECH | Inventor | Manufacturer

8 个月

In assessing the impact of trillions of parameters on LLM effectiveness, historical data suggests that while larger models can enhance performance in certain tasks, they also come with trade-offs in terms of computational resources and scalability. However, testing a focused data-based LLM offers insights into optimizing the balance between effectiveness, performance, and cost. Considering advancements in parallel processing and distributed computing, have you explored techniques like model distillation or pruning to mitigate resource constraints while maintaining model efficacy?

回复

要查看或添加评论,请登录

Venkat Ramakrishnan的更多文章

  • On RAGs and Riches

    On RAGs and Riches

    Back in 2018, when I did a talk at ThoughtWorks on NLP, there was an euphoria on the state of chatbots. There was even…

  • The System Testing Of AI

    The System Testing Of AI

    When we test systems, we don't stop with just testing of functionality of modules, or integration testing of the…

  • A bit about hallucinations

    A bit about hallucinations

    While LLMs are hot, their hallucinations are stark. For a casual user of the LLMs, they might seem to be minor mistakes…

  • At Wit's End On LLM performance?

    At Wit's End On LLM performance?

    Nowadays LLMs' performance is a daily topic! Me, like you, go awestruck looking at those magical numbers when an…

  • The Curious Case Of Software Naming

    The Curious Case Of Software Naming

    You all call me 'Venkat', and I'm okay with that! To be honest, there are boatloads of 'Venkat Ramakrishnan's out…

  • Prevention Is Better Than Cure

    Prevention Is Better Than Cure

    These past forty-five days or so saw the rise of voices of cybersecurity professionals from various capacities towards…

    2 条评论
  • Integration Nightmare: The Case Of Super-flexible e-commerce platforms

    Integration Nightmare: The Case Of Super-flexible e-commerce platforms

    Freedom comes at a cost, which is not devoting ourselves to what we know well and accustomed to. This is especially…

  • Rocket Science: An Emerging Quality and Testing Opportunity

    Rocket Science: An Emerging Quality and Testing Opportunity

    A few months back, I had attended a startup enclave in Bengaluru in which I met a variety of entrepreneurs, some…

  • Verify, Then Trust

    Verify, Then Trust

    These are strange times that we live in wherein we cannot trust implicitly without verifying. There were times when we…

  • To Bell The Generative AI Cat

    To Bell The Generative AI Cat

    The latest survey by Weber Shandwick shows that Generative AI is a top priority for CEOs, with 51% marking it as a top…

社区洞察

其他会员也浏览了