LLM benchmarking for large language models improvement

Toloka is developing several benchmarks tailored for sectors with a high need for precision assessment to address the rising demand for precise LLM evaluation. In collaboration with top universities, we developed Beemo, an advanced benchmark for detecting AI-generated text. It’s a vital tool for identifying and evaluating synthetic content that can be used in training datasets. Additionally, Toloka assessed LLM performance across natural sciences, creating a detailed benchmark spanning ten subject areas to evaluate model accuracy and reliability in complex scientific contexts. 

Read more in Toloka’s blog posts on Beemo and natural sciences benchmark.

As the capabilities of large language models (LLMs) expand, the need for robust, nuanced, and domain-specific benchmarks is more critical than ever. Benchmarks are crucial in measuring these models' performance, safety, and alignment with intended tasks. They help developers and researchers assess whether LLMs meet the demands of real-world applications. This article will delve into LLM benchmarking, benchmark datasets, and the pioneering datasets Toloka is developing to address specific challenges in AI.

What is LLM benchmarking?

Benchmarking of an LLM is the process of evaluating large language models and comparing performance metrics to measure their performance in specific tasks or domains. Evaluating LLMs involves metrics or scorers to assess performance. They offer a standardized way to compare model outputs. 

This assessment also involves using carefully designed datasets, which are also called benchmarks, and contain tasks intended to highlight a model’s strengths and weaknesses. These benchmarks (metrics and datasets) help developers understand how well their LLMs perform in areas like language comprehension, factual accuracy, reasoning, and even specialized fields like mathematics or natural sciences. 

Why is it essential for LLM evaluation?

Large language model evaluation through benchmarks ensures that these models are accurate, dependable, and well-suited to users' and industries' needs. It’s a key part of developing helpful and responsible AI. Evaluating LLMs is essential because it clearly shows how well these models perform, where they may fall short, and how they can be improved.

Uncovering weaknesses and errors

Benchmarks help identify areas where LLMs may produce inaccurate or misleading information, such as hallucinations that are factually incorrect model responses. They can also show where the model struggles with complex reasoning or inconsistencies in specialized knowledge domains. These issues can go unnoticed without systematic evaluation, potentially leading to trust issues or unintended consequences in future applications.

Improving models where it matters

The evaluation highlights specific weaknesses, making it easier to make improvements. Based on these findings, developers can fine-tune models to boost accuracy in areas that need it most. For instance, if a model has trouble with scientific concepts or math reasoning, targeted training data can help fill those gaps.

Ensuring models are useful across various domains

General-purpose LLMs often underperform in specialized areas like medicine, law, or technical. Specialized benchmarks help developers check whether their models are ready for real-world use in these domains, making them safer and more reliable for professionals who need accuracy.

Building responsible and ethical AI

Reliable benchmarking helps ensure that models follow ethical guidelines, such as detecting bias, keeping content authentic, and identifying AI-generated text when needed. This helps build models that are accurate and aligned with ethical standards, ultimately building trust.

Adapting models to changing Information

Benchmarks help keep LLMs relevant in rapidly changing fields. For instance, models used in medicine or finance need constant updating to stay accurate with new research or policies. Regular benchmarking highlights areas where models might rely on outdated information so developers can keep them current. 

How LLM benchmarks work

LLM benchmarks offer a structured way to evaluate and compare language models on various tasks. Benchmarks work by defining specific tasks or skill areas that are relevant to understanding the capabilities of LLMs. Here are some core tasks benchmarks commonly assess:

Language generation and conversation. Measures the model’s ability to generate coherent, contextually relevant responses to prompts or questions, essential for tasks like creative writing and chat-based dialogue;

Understanding and answering questions. Evaluates the model’s skill in interpreting text and producing accurate, relevant answers. This is crucial for information retrieval tasks and question-answering systems;

Translation. Tests the model's ability to translate between languages with fluency and accuracy, allowing models to handle multilingual tasks;

Logical reasoning and common sense. Involves tasks that test the model’s ability to use logic, reasoning, and everyday knowledge to solve problems. This can include both inductive and deductive reasoning challenges.

Standardized tests. Benchmarks may include tasks modeled after human exams like the SAT, ACT, and other standardized assessments, offering insight into how well an LLM mimics human learning and reasoning abilities.

Code generation. Assesses the model’s capability to understand programming-related tasks and generate functional code. This task area is essential for evaluating models aimed at supporting software development.

Evaluation strategies

There are two basic strategies for LLM evaluation: offline and online evaluation. The offline approach involves testing LLM performance in a controlled setting before deployment. It’s a reliable way to identify and address weaknesses without exposing end-users to errors.

Once the model is live, online evaluation helps monitor its ongoing performance during real-world use. This feedback loop ensures the model meets quality standards and adapts to evolving user needs.

Benchmarking metrics

Benchmarking metrics are the scoring criteria used to grade the model’s performance. They tell us how well a model performs on specific tasks and where it might need improvement. Because human language is very complex and diverse, many approaches and metrics exist for evaluating LLMs. These metrics range from straightforward statistics-based scores to more specialized scorers, even involving LLMs that can evaluate other models (LLM-assisted evaluation). 

Types of metrics

Although there is no single classification of LLM evaluation approaches, the following types of metrics can be singled out.

Basic statistical metrics

Some simpler metrics focus on direct comparisons, such as counting common word patterns or sequences. BLEU and ROUGE are examples often used for tasks like translation and summarization, where matching specific phrases and structures matters.

Contextual or domain-specific scorers

Other metrics, like METEOR or BLEURT, consider context, tone, and meaning, allowing for a more human-like evaluation beyond word-for-word matching. Domain-specific scorers measure performance in specialized fields like code generation or complex reasoning, focusing on whether a model can fulfill specific requirements accurately, as seen in benchmarks like HumanEval for coding.

LLM-Assisted Evaluation

In some cases, newer LLMs act as judges to evaluate the outputs of other models, simulating human judgment. Advanced models like GPT-4 can assess criteria such as relevance or helpfulness, which is especially useful for complex tasks like dialogue or multi-turn questions.

The importance of golden datasets

Behind almost all these scorers lies a golden dataset with high-quality examples and accurate answers for various tasks. Much like a human learns and gets certified in a field, these datasets guide LLMs during training, fine-tuning, and benchmarking, ensuring they have a well-defined reference to meet real-world standards. 

The following are some widely used LLM benchmarks or metrics and what they measure.

Common LLM benchmarks

• Accuracy or precision measures how often the model's answers are correct; 

• Recall shows how well the model captures all the relevant instances, or in other words, it tells us how good the model is at finding everything it's supposed to find;

• F1 Score is a balance of precision and recall, giving a single number showing how well the model finds and correctly identifies the positive answers. It is beneficial when there is a mix of right and wrong answers;

• Perplexity is used for language tasks; perplexity checks how well the model predicts the next word in a sentence. Lower perplexity means it’s better at making text sound natural;

• Exact Match (EM) is a metric used to measure how often a model’s prediction exactly matches the correct answer;

• BLEU (Bilingual Evaluation Understudy) is an algorithm that scores the similarity between a model-generated text and one or more reference texts, typically using n-gram precision as its core evaluation method.

• ROUGE (Recall-Oriented Understudy for Gisting Evaluation) is a set of metrics and software often used to evaluate summarization tasks by measuring the overlap between model-generated and reference summaries. It primarily evaluates recall, meaning how much of the reference text appears in the model’s output;

• METEOR (Metric for Evaluation of Translation with Explicit Ordering) is an automatic metric that uses exact matches, synonyms, stemming, and word order to measure similarity, allowing for more varied, human-like phrasing in the model’s output. It improves by calculating both precision (n-gram matches) and recall (n-gram overlaps) between the model's output and reference text;

• BLEURT (Bilingual Evaluation Understudy with Representations from Transformers) is an automatic metric that leverages pre-trained Transformer models such as BERT to assess the similarity between generated and reference texts.

Benchmarking datasets

Benchmarking datasets are data collections like questions or tasks used to test and train LLMs on specific abilities or knowledge areas. Unlike generic training datasets, which help the model learn, benchmarking datasets are designed to probe the model’s strengths and weaknesses in targeted areas.

They often cover a range of tasks, which helps test models in different scenarios. These might include everything from simple classification tasks to more complex activities like logical reasoning or creative generation.

Some widely used datasets to benchmark LLM 

• MMLU (Massive Multitask Language Understanding) is a benchmark that tests LLMs’ ability to perform well across a large variety of tasks simultaneously, reflecting how models handle general knowledge and domain-specific questions. MMLU includes questions from 57 domains, starting from basic subjects like math and history to more specialized fields such as law or medicine. It’s structured as a multiple-choice dataset, where each question has four possible answers;

• HELM (Holistic Evaluation of Language Models) evaluates models from multiple angles, providing a holistic view of performance. It uses established datasets to assess accuracy and performance across various tasks and integrates qualitative reviews to capture subtleties in the model's responses. HELM also conducts error analysis to pinpoint specific areas where a model may struggle;

• GLUE (General Language Understanding Evaluation) is a comprehensive NLP benchmark comprising nine key tasks, each designed to test different aspects of language understanding, such as sentence classification, sentiment analysis, and question answering. For each task, GLUE provides a training set to train models, a development set for fine-tuning and adjusting model parameters, and an evaluation set to assess final performance;

• HellaSwag is a benchmark dataset designed to evaluate commonsense reasoning and contextual understanding in LLMs. Unlike standard multiple-choice tasks, HellaSwag presents models with a context and several possible continuations, but only one of these continuations is truly logical based on the given context;

• MATH is a specialized benchmark dataset consisting of 12,500 challenging mathematics problems, primarily sourced from high-school level and competition math. The dataset is designed to assess and advance the problem-solving and mathematical reasoning abilities of LLMs;

• HumanEval is a benchmark dataset that evaluates the ability of large language models to generate functional code. It consists of 164 programming tasks, each paired with a problem description and a function signature but without the complete solution.

Toloka’s approach to benchmarking

Toloka's approach to LLM benchmarking centers on evaluating models in areas where they are most vulnerable. Recent research by our team identified critical domains like AI-generated content detection, university-level math, and science-related domains as priority areas for improvement.

Specialized targeted benchmarks were developed at Toloka to help LLM developers assess their models' performance in these vulnerable domains. This initiative has led to the creation of three benchmarks, each designed to uncover specific performance gaps in language models.

Toloka utilizes an AI benchmarking evaluation method that measures models' performance on specific tasks using tailored datasets to address the challenges facing popular language models. This approach allows LLM producers to identify gaps in their models' capabilities and assess whether additional fine-tuning with supplementary data may be necessary to improve accuracy and alignment.

A tool for reliable AI-generated text detection

One essential benchmark Toloka developed is Beemo, a tool for detecting AI-generated text, which is crucial for ensuring data quality in LLM training. Beemo is unique in that it combines AI-generated and human-edited text within one dataset. 

University-level math benchmark

Despite advancements in LLMs, their reasoning abilities remain limited, particularly for complex subjects like mathematics. While many existing benchmarks assess basic math skills, they primarily focus on school-level problems, leaving a considerable gap in evaluating how LLMs handle more advanced, university-level questions. 

To address this, Toloka developed the largest benchmark of its kind, featuring 20,000 real-world university-level math problems. This dataset allows for a comprehensive evaluation of LLMs’ mathematical reasoning, logical progression, and problem-solving skills at an advanced academic level. 

Benchmarking for natural science LLMs

In response to the need for accurate assessments of LLMs in science-related areas, Toloka developed a specialized dataset that spans multiple scientific domains. Many existing datasets lack the complexity to evaluate models on the challenging questions scientists encounter in their work. 

To fill this gap, Toloka collaborated with domain experts who were actively researching fields like high-energy physics, immunology, and cell biology. 

How to use benchmarking datasets

To get the most out of benchmarking datasets, it's important to choose tasks and metrics that align well with what the model is expected to accomplish. For example, if the model is designed for customer support, it should be evaluated on datasets that reflect typical customer inquiries and responses. 

The metrics should measure accuracy and criteria like the relevance and coherence of the responses. A mix of benchmarks that tap into different skills LLM might need in the real world are usually employed to assess a language model's capabilities comprehensively.

Benchmark datasets typically consist of structured questions or prompts and gold-standard answers across a wide range of topics and complexity levels. First, the model runs through these tasks, and its responses are recorded for comparison against the benchmark answers.

After the model completes the tasks, each benchmark provides specific scoring metrics designed for that type of content, such as accuracy, coherence, or relevance. These scores clearly show the model's performance across different skills or knowledge areas. By comparing results with other models, these benchmarks help to spot strengths, identify improvement areas, and guide further tuning. 

Such datasets help systematically evaluate a model's performance and highlight its strengths and areas needing improvement. Developers can gain valuable insights into refining their models for better performance by carefully choosing tasks that align with their goals and looking closely at the results. 

This thoughtful approach enhances the model’s capabilities and builds trust in the technology as it adapts to meet user needs. Ultimately, investing time in benchmarking can lead to more innovative and practical applications of AI.