Large language model

Using a modification of byte-pair encoding, in the first step, all unique characters (including blanks and punctuation marks) are treated as an initial set of n-grams (i.e. initial set of uni-grams). Successively the most frequent pair of adjacent characters is merged into a bi-gram and all instances of the pair are replaced by it. All occurrences of adjacent pairs of (previously merged) n-grams that most frequently occur together are then again merged into even lengthier n-gram repeatedly until a vocabulary of prescribed size is obtained (in case of GPT-3, the size is 50257).[6] Token vocabulary consists of integers, spanning from zero up to the size of the token vocabulary. New words can always be interpreted as combinations of the tokens and the initial-set uni-grams.[7]

A token vocabulary based on the frequencies extracted from mainly English corpora uses as few tokens as possible for an average English word. An average word in another language encoded by such an English-optimized tokenizer is however split into suboptimal amount of tokens.

tokenizer: texts -> series of numerical "tokens" may be split into:

Probabilistic tokenization also compresses the datasets, which is the reason for using the byte pair encoding algorithm as a tokenizer. Because LLMs generally require input to be an array that is not jagged, the shorter texts must be "padded" until they match the length of the longest one. How many tokens are, on average, needed per word depends on the language of the dataset.[8][9]

Removal of toxic passages from the dataset, discarding low-quality data, and de-duplication are examples of dataset cleaning.[10] Resulting, cleaned (high-quality) datasets contain up to 17 trillion words in 2022, raising from 985 million words, used in 2018 for GPT-1,[11] and 3.3 billion words, used for BERT.[12] The future data is, however, expected to be increasingly "contaminated" by LLM-generated contents themselves.[13]

Reinforcement learning from human feedback (RLHF) through algorithms, such as proximal policy optimization, is used to further fine-tune a model based on a dataset of human preferences.[14]

Using "self-instruct" approaches, LLMs have been able to bootstrap correct responses, replacing any naive responses, starting from human-generated corrections of a few cases. For example, in the instruction "Write an essay about the main themes represented in Hamlet," an initial naive completion might be 'If you submit the essay after March 17, your grade will be reduced by 10% for each day of delay," based on the frequency of this textual sequence in the corpus.[15]

The largest LLM may be too expensive to train and use directly. For such models, mixture of experts (MoE) can be applied, a line of research pursued by Google researchers since 2017 to train models reaching up to 1 billion parameters.[16][17][18]

Most results previously achievable only by (costly) fine-tuning, can be achieved through prompt engineering, although limited to the scope of a single conversation (more precisely, limited to the scope of a context window).[19]

When each head calculates, according to its own criteria, how much other tokens are relevant for the "it_" token, note that the second attention head, represented by the second column, is focusing most on the first two rows, i.e. the tokens "The" and "animal", while the third column is focusing most on the bottom two rows, i.e. on "tired", which has been tokenized into two tokens.[20]

In order to find out which tokens are relevant to each other within the scope of the context window, the attention mechanism calculates "soft" weights for each token, more precisely for its embedding, by using multiple attention heads, each with its own "relevance" for calculating its own soft weights. For example, the small (i.e. 117M parameter sized) GPT-2 model, has had twelve attention heads and a context window of only 1k token.[21] In its medium version it has 345M parameters and contains 24 layers, each with 12 attention heads. For the training with gradient descent a batch size of 512 was utilized.[7]

The largest models can have a context window sized up to 32k (for example, GPT-4; while GPT-3.5 has a context window sized from 4k to 16k, and legacy GPT-3 has had 2k sized context window).[22]

Length of a conversation that the model can take into account when generating its next answer is limited by the size of a context window, as well. If the length of a conversation, for example with Chat-GPT, is longer than its context window, only the parts inside the context window are taken into account when generating the next answer, or the model needs to apply some algorithm to summarize the too distant parts of conversation.

The shortcomings of making a context window larger include higher computational cost and possibly diluting the focus on local context, while making it smaller can cause a model to miss an important long-range dependency. Balancing them are a matter of experimentation and domain-specific considerations.

A model may be pre-trained either to predict how the segment continues, or what is missing in the segment, given a segment from its training dataset.[23] It can be either

• autoregressive (i.e. predicting how the segment continues, the way GPTs do it): for example given a segment "I like to eat", the model predicts "ice cream", or
• "masked" (i.e. filling in the parts missing from the segment, the way "BERT"[12] does it): for example, given a segment "I like to [__] [__] cream", the model predicts that "eat" and "ice" are missing.

Models may be trained on auxiliary tasks which test their understanding of the data distribution, such as Next Sentence Prediction (NSP), in which pairs of sentences are presented and the model must predict whether they appear consecutively in the training corpus.[12] During training, regularization loss is also used to stabilize training. However regularization loss is usually not used during testing and evaluation.

The following four hyper-parameters characterize a LLM:

• cost of (pre-)training (${\displaystyle C}$),
• size of the artificial neural network itself, such as number of parameters ${\displaystyle N}$ (i.e. amount of neurons in its layers, amount of weights between them and biases),
• size of its (pre-)training dataset (i.e. number of tokens in corpus, ${\displaystyle D}$),
• performance after (pre-)training.

They are related by simple statistical laws, called "scaling laws". One particular scaling law ("Chinchilla scaling") for LLM autoregressively trained for one epoch, with a log-log learning rate schedule, states that:[60]

$${\displaystyle {\begin{cases}C=C_{0}ND\\L={\frac {A}{N^{\alpha }}}+{\frac {B}{D^{\beta }}}+L_{0}\end{cases}}}$$

where the variables are

• ${\displaystyle C}$ is the cost of training the model, in FLOPs.
• ${\displaystyle N}$ is the number of parameters in the model.
• ${\displaystyle D}$ is the number of tokens in the training set.
• ${\displaystyle L}$ is the average negative log-likelihood loss per token (nats/token), achieved by the trained LLM on the test dataset.

and the statistical hyper-parameters are

• ${\displaystyle C_{0}=6}$, meaning that it costs 6 FLOPs per parameter to train on one token. Note that training cost is much higher than inference cost, where it costs 1 to 2 FLOPs per parameter to infer on one token.[28]
• ${\displaystyle \alpha =0.34,\beta =0.28,A=406.4,B=410.7,L_{0}=1.69}$

At point(s) referred to as breaks,[61] the lines change their slopes, appearing on a log-log plot as a series of linear segments connected by arcs.

When one subtracts out from the y-axis the best performance that can be achieved even with infinite scaling of the x-axis quantity, large models' performance, measured on various tasks, seems to be a linear extrapolation of other (smaller-sized and medium-sized) models' performance on a log-log plot. However, sometimes the line's slope transitions from one slope to another at point(s) referred to as break(s)[61] in downstream scaling laws, appearing as a series of linear segments connected by arcs; it seems that larger models acquire "emergent abilities" at this point(s).[19][62] These abilities are discovered rather than programmed-in or designed, in some cases only after the LLM has been publicly deployed.[3]

The most intriguing among emergent abilities is in-context learning from example demonstrations.[63] In-context learning is involved in tasks, such as:

• reported arithmetics, decoding the International Phonetic Alphabet, unscrambling a word's letters, disambiguate word in context,[19][64][65] converting spatial words, cardinal directions (for example, replying "northeast" upon [0, 0, 1; 0, 0, 0; 0, 0, 0]), color terms represented in text.[66]
• chain-of-thought prompting: Model outputs are improved by chain-of-thought prompting only when model size exceeds 62B. Smaller models perform better when prompted to answer immediately, without chain of thought.[67]
• identifying offensive content in paragraphs of Hinglish (a combination of Hindi and English), and generating a similar English equivalent of Kiswahili proverbs.[68]

Schaeffer et. al. argue that the emergent abilities are not unpredictably acquired, but predictably acquired according to a smooth scaling law. The authors considered a toy statistical model of an LLM solving multiple-choice questions, and showed that this statistical model, modified to account for other types of tasks, applies to these tasks as well.[69]

Let ${\displaystyle x}$ be the number of parameter count, and ${\displaystyle y}$ be the performance of the model.

NLP researchers were evenly split when asked, in a 2022 survey, whether (untuned) LLMs "could (ever) understand natural language in some nontrivial sense".[74] Proponents of "LLM understanding" believe that some LLM abilities, such as mathematical reasoning, imply an ability to "understand" certain concepts. A Microsoft team argued in 2023 that GPT-4 "can solve novel and difficult tasks that span mathematics, coding, vision, medicine, law, psychology and more" and that GPT-4 "could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence system": "Can one reasonably say that a system that passes exams for software engineering candidates is not really intelligent?"[75][76] Some researchers characterize LLMs as "alien intelligence".[77][78] For example, Conjecture CEO Connor Leahy considers untuned LLMs to be like inscrutable alien "Shoggoths", and believes that RLHF tuning creates a "smiling facade" obscuring the inner workings of the LLM: "If you don't push it too far, the smiley face stays on. But then you give it [an unexpected] prompt, and suddenly you see this massive underbelly of insanity, of weird thought processes and clearly non-human understanding."[79][80]

In contrast, some proponents of the "LLMs lack understanding" school believe that existing LLMs are "simply remixing and recombining existing writing",[78] or point to the deficits existing LLMs continue to have in prediction skills, reasoning skills, agency, and explainability.[74] For example, GPT-4 has natural deficits in planning and in real-time learning.[76] Generative LLMs have been observed to confidently assert claims of fact which do not seem to be justified by their training data, a phenomenon which has been termed "hallucination".[81] Specifically, hallucinations in the context of LLMs correspond to the generation of text or responses that seem syntactically sound, fluent, and natural but are factually incorrect, nonsensical, or unfaithful to the provided source input.[82] Neuroscientist Terrence Sejnowski has argued that "The diverging opinions of experts on the intelligence of LLMs suggests that our old ideas based on natural intelligence are inadequate".[74]

The matter of LLM's exhibiting intelligence or understanding [74] has foundations in the study of language as a model of Cognition in the field of Cognitive linguistics. The American Linguist George Lakoff presented Neural Theory of Language (NTL) as a computational basis for using language as a model of learning tasks and understanding.[83] In his 2014 book titled The Language Myth: Why Language Is Not An Instinct, British cognitive linguist and digital communication technologist Vyvyan Evans maps out the role of probabilistic context-free grammar (PCFG) in enabling NLP to model cognitive patterns.

The most commonly used measure of a language model's performance is its perplexity on a given text corpus. Perplexity is a measure of how well a model is able to predict the contents of a dataset; the higher the likelihood the model assigns to the dataset, the lower the perplexity. Mathematically, perplexity is defined as the exponential of the average negative log likelihood per token:

$${\displaystyle \log({\text{Perplexity}})=-{\frac {1}{N}}\sum _{i=1}^{N}\log(Pr({\text{token}}_{i}|{\text{context for token}}_{i}))}$$

here ${\displaystyle N}$ is the number of tokens in the text corpus, and "context for token ${\displaystyle i}$" depends on the specific type of LLM used. If the LLM is autoregressive, then "context for token ${\displaystyle i}$" is the segment of text appearing before token ${\displaystyle i}$. If the LLM is masked, then "context for token ${\displaystyle i}$" is the segment of text surrounding token ${\displaystyle i}$.

Because language models may overfit to their training data, models are usually evaluated by their perplexity on a test set of unseen data.[12] This presents particular challenges for the evaluation of large language models. As they are trained on increasingly large corpora of text largely scraped from the web, it becomes increasingly likely that models' training data inadvertently includes portions of any given test set.[4]

A large number of testing datasets and benchmarks have also been developed to evaluate the capabilities of language models on more specific downstream tasks. Tests may be designed to evaluate a variety of capabilities, including general knowledge, commonsense reasoning, and mathematical problem-solving.

One broad category of evaluation dataset is question answering datasets, consisting of pairs of questions and correct answers, for example, ("Have the San Jose Sharks won the Stanley Cup?", "No").[84] A question answering task is considered "open book" if the model's prompt includes text from which the expected answer can be derived (for example, the previous question could be adjoined with some text which includes the sentence "The Sharks have advanced to the Stanley Cup finals once, losing to the Pittsburgh Penguins in 2016."[84]). Otherwise, the task is considered "closed book", and the model must draw on knowledge retained during training.[85] Some examples of commonly used question answering datasets include TruthfulQA, Web Questions, TriviaQA, and SQuAD.[85]

Evaluation datasets may also take the form of text completion, having the model select the most likely word or sentence to complete a prompt, for example: "Alice was friends with Bob. Alice went to visit her friend, ____".[4]

Some composite benchmarks have also been developed which combine a diversity of different evaluation datasets and tasks. Examples include GLUE, SuperGLUE, MMLU, BIG-bench, and HELM.[86][85]

It was previously standard to report results on a heldout portion of an evaluation dataset after doing supervised fine-tuning on the remainder. It is now more common to evaluate a pre-trained model directly through prompting techniques, though researchers vary in the details of how they formulate prompts for particular tasks, particularly with respect to how many examples of solved tasks are adjoined to the prompt (i.e. the value of n in n-shot prompting).

Because of the rapid pace of improvement of large language models, evaluation benchmarks have suffered from short lifespans, with state of the art models quickly "saturating" existing benchmarks, exceeding the performance of human annotators, leading to efforts to replace or augment the benchmark with more challenging tasks.[87] In addition, there are cases of "shortcut learning" wherein AIs sometimes "cheat" on multiple-choice tests by using statistical correlations in superficial test question wording in order to guess the correct responses, without necessarily understanding the actual question being asked.[74]

Some datasets have been constructed adversarially, focusing on particular problems on which extant language models seem to have unusually poor performance compared to humans. One example is the TruthfulQA dataset, a question answering dataset consisting of 817 questions which language models are susceptible to answering incorrectly by mimicking falsehoods to which they were repeatedly exposed during training. For example, an LLM may answer "No" to the question "Can you teach an old dog new tricks?" because of its exposure to the English idiom you can't teach an old dog new tricks, even though this is not literally true.[88]

Another example of an adversarial evaluation dataset is Swag and its successor, HellaSwag, collections of problems in which one of multiple options must be selected to complete a text passage. The incorrect completions were generated by sampling from a language model and filtering with a set of classifiers. The resulting problems are trivial for humans but at the time the datasets were created state of the art language models had poor accuracy on them. For example:

We see a fitness center sign. We then see a man talking to the camera and sitting and laying on a exercise ball. The man...
a) demonstrates how to increase efficient exercise work by running up and down balls.
b) moves all his arms and legs and builds up a lot of muscle.
c) then plays the ball and we see a graphics and hedge trimming demonstration.
d) performs sits ups while on the ball and talking.[89]

BERT selects b) as the most likely completion, though the correct answer is d).[89]