Fine-tuning LLMs for Enterprise

Introduction

The popularity of ChatGPT has introduced the world to large language models (LLMs) and their extraordinary abilities in performing natural language tasks.

According to Accenture, such tasks account for 62% of office workers’ time, and 65% of that could be made more productive through using LLMs to automate or augment Enterprise working practises thereby boosting productivity, innovation, and customer engagement.

To give some examples, LLMs could be integrated into Customer Services to handle product queries, thereby improving response times and customer satisfaction. Equally, LLMs could assist in drafting articles, scripts, or promotional materials, or be used by analysts for summarising vast amounts of information, or gauging market sentiment by analysing customer reviews and feedback.

Whilst potentially disruptive and likely to lead to some job losses (by the mid-2030s, up to 30% of jobs could be automated), this disruption and new way of working is also forecast to grow global revenues by 9%.

It’s perhaps not surprising then that Enterprise executives are showing a keen interest in LLMs and the role they could play in their organisations’ strategies over the next 3 to 5 years.

Large language models such as OpenAI’s GPT-4 or GPT-3.5 (upon which ChatGPT is based) or open source alternatives such as Meta’s recently launched Llama2, are what’s known as foundation models.

Such models are pre-trained on a massive amount of textual data and then tuned through a process of alignment to be performant across a broad range of natural language tasks. Crucially though, their knowledge is limited by the extent of the data they were trained on, and their behaviour is dictated by the approach and objectives employed during the alignment phase.

To put it bluntly, a foundational LLM, whilst exhibiting a dazzling array of natural language skills, is less adept at generating legal documents or summarising medical information, and may be inadequate for those Customer Support applications requiring more empathy, and will certainly lack detailed knowledge on a particular product or business.

To be truly useful therefore, LLMs need to be adapted to the domain and particular use cases where they’ll be employed. This article explores the options for tailoring LLMs for Enterprise use; a follow-up article here, looks further forward into how LLMs can be enhanced to improve their knowledge, understanding, reasoning and ability to scale.

Domain-specific pre-training

One approach would be to collect domain-specific data and train a new model.

However, pre-training your own LLM from scratch is not easy, requiring massive amounts of data, lots of expensive compute hours for training the model, and a dedicated team working on it for weeks or even months. As a result, very few organisations choose this path, although notable examples include BloombergGPT (finance) and Med-PaLM 2 (medicine) and Nvidia have recently launched the NeMo framework to lend a helping hand.

Nonetheless, training a dedicated model is a serious undertaking and only open to those with the necessary resources. For everyone else, an alternate (and arguably easier) approach is to start with an existing foundational model such as GPT-3.5 and fine-tune from there.

Fine-tuning

As a form of transfer learning, fine-tuning adapts the parameters within a foundational model to better perform particular tasks.

Guidance from OpenAI for gpt-3.5-turbo indicates that 50–100 well-crafted examples is usually sufficient to fine-tune a model, although the amount will ultimately depend on the use case.

In comparison to domain-specific pre-trained models which require lots of resource, fine-tuning a foundational model requires less data, costs less, and can be completed in days, putting it well within the reach of many companies.

But it’s not without its drawbacks…

A common misconception is that fine-tuning enables the model to acquire new information, but in reality it only teaches it to perform better within particular tasks, a goal which can also be achieved through careful prompting as we’ll see later.

Fine-tuning also won’t prevent hallucinations that undermine the reliability and trustworthiness of the model’s output; and there is always a risk of introducing biases or inaccuracies into the model via the examples chosen, or inadvertently training it with sensitive information which subsequently leaks out (hence consideration should be given to using synthetic data).

Where support is required for a diverse set of tasks or edge cases within a given domain, relying on fine-tuning alone might result in a model that is too generic, performing poorly against each subtask. In such a situation, individual models may need to be created for each task and updated frequently to stay current and relevant as new knowledge becomes available, hence becoming a resource-intensive and cumbersome endeavour.

Fortunately, there are other techniques that can be employed, either in concert with or replacing fine-tuning entirely — prompt engineering.

Few-shot prompting

Irrespective of how a language model has been pre-trained and whether or not it’s been fine-tuned, the usefulness of its output is directly related to the quality of the prompt it receives. As so aptly put by OpenAI, “GPTs can’t read your mind”.

Although models can perform relatively well when prompted in a zero-shot manner (i.e., comprising just the task description and any input data), they can also be inconsistent, and may try to answer a question by regurgitating random facts or making something up from their training data (i.e., hallucinating) — they might know how words relate statistically, but they don’t know what they mean.

Output can be improved by supplementing the prompt with one or more input/output examples (few-shot) that provide context to the instruction as well as guidance on desired format, style of response and length; this is known as in-context learning (ICL); see below (Source: Daniel Bashir, The Gradient):

The order in which examples are provided can impact a model’s performance, as can the format used. Diversity is also incredibly important, models prompted with a diverse set of examples tending to perform better (although only the larger foundational models such as GPT-4 cope well with examples that diverge too far from what the model was originally pre-trained with).

Retrieval Augmented Generation

A good way of achieving this diversity is to retrieve task-specific examples from domain-specific knowledge sources using frameworks such as LlamaIndex, LangChain, HoneyHive, Lamini or Microsoft’s LLM-AUGMENTER.

Commonly referred to as Retrieval Augmented Generation, this approach ensures that the model has access to the most current and reliable domain-specific facts (rather than the static corpus it was pre-trained with), and users have visibility of the model’s sources thereby enabling its responses to be checked for accuracy.

As so aptly put by IBM Research, “It’s the difference between an open-book and a closed-book exam”, and hence it’s not surprising that LLMs perform much better when provided with external information sources to draw upon.

A straightforward way of implementing the RAG method is via a keyword search to retrieve relevant text chunks from external documentation, but a better approach is to use embeddings.

Put simply, embedding is a process by which the text is tokenised and passed through the LLM to create a numerical representation of the semantic meaning of the words and phrases within the text, and this representation is then be stored in a vector database (such as Pinecone, Weaviate or Chroma).

Upon receiving a query, the RAG system conducts a vector search of the database based on an embedding of the user query, retrieves relevant text chunks based on similarity and appends them to the prompt for feeding into the LLM:

Care though is needed to not overload the prompt with too much information as any increase in the prompt size directly increases the compute, time and cost for the LLM to derive an output (computation increasing quadratically with input length), and also risks exceeding the foundation model’s max prompt window size (and especially so in the case of open source models which typically have much smaller windows).

Whilst providing additional context and task-specific data should reduce the instances of hallucinations, LLMs still struggle with complex arithmetic, common sense, or symbolic reasoning, hence attention is also needed to the way the LLM is instructed to perform the task, an approach known as instruction prompting.

Instruction prompting

Chain of Thought (CoT) is one such technique, explored by Google and OpenAIamongst others, in which the model is directly instructed to follow smaller, intermediate steps towards deriving the final answer. Extending the prompt instruction with a phrase as simple as “Let’s consider step by step…” can have a surprising effect in helping the model to break down the task into steps rather than jumping in with a quick, and often incorrect, answer.

Self-ask is a similar approach in which the model is asked to generate and then answer sub-questions about the input query first (and with the option of farming out these sub-questions to Google Search to retrieve up-to-date answers), before then using this knowledge to compile the final answer (essentially a combination of CoT and RAG).

Yet another technique, Tree of Thoughts (ToT) is similar in generating a solution based on a sequence of individual thoughts, but goes further by allowing multiple reasoning paths to be considered simultaneously (forming a tree of potential thoughts) and exploring each in turn before settling on a final answer.

Source: Yao, Shunyu, et al. “Tree of thoughts: Deliberate problem solving with large language models.” arXiv preprint arXiv:2305.10601 (2023)

Whilst proven to be effective, these various instruction prompting techniques take a linear approach that progresses from one thought to the next. Humans think a little differently, following and sometimes combining insights from different chains of thought to arrive at the final answer. This reasoning process can be modelled as a graph structure and forms yet another area of research.

A final technique, which might seem even more peculiar than asking the model to take a stepwise approach (CoT and ToT) is to assign it a “role” or persona within the prompt such as “You are a famous and brilliant mathematician”. Whilst this role based prompting may seem bizarre, it’s actually providing the model with additional context to better understand the question, and has been found surprisingly to produce better answers.

Options & considerations

The previous sections have identified a range of techniques that can be employed to contextualise an LLM to Enterprise tasks, but which should you choose?

The first step is to choose whether to generate your own domain pre-trained model, fine-tune an existing foundational model, or simply rely on prompting at runtime:

There’s more discussion later on around some of the criteria to consider when selecting which foundational model to use…

Fine-tuning may at first seem the most logical path, but requires a careful investment of time and effort, hence sticking with a foundational model and experimenting with the different prompting techniques is often the best place to start, a sentiment echoed by OpenAI in their guidance for GPT.

Choice of which techniques to try will be dependent on the nature of the task:

Good results can often be achieved by employing different prompting techniques in combination:

• Detailed instructions (instruction prompting) — especially where the task involves complex reasoning
• Carefully chosen set of examples (few-shot learning) — to demonstrate the tone, format and length of output that is required
• Supplementary information (in-context learning, RAG & embeddings) — retrieved from domain-specific knowledge sources to provide more context

It’s also about balance — few-shot learning typically consumes a lot of tokens which can be problematic given the limited window size of many LLMs. So rather than guiding the model in terms of desired behaviour via a long set of examples, this can be offset by incorporating a more precise, textual description of what’s required via instruction prompting.

Prompt window size can also be a limitation in domains such as medical and legal which are more likely to require large amounts of information to be provided in the prompt; for instance most research papers (~5–8k tokens) would exceed the window size of the base GPT-3.5 model as well as many of the open source LLMs which typically only support up to 2,000 tokens (~1,500 words).

Choosing a different LLM with a larger window is certainly an option (GPT-4 can extend to 32k tokens), but as mentioned earlier will quadratically increase the amount of compute, time and cost needed to complete the task, hence in such applications it may be more appropriate to fine-tune the LLM, despite the initial outlay.

Model size is yet another factor that needs to be considered. Pre-training a domain-specific LLM, or fine-tuning a small foundational model (such as GPT-3.5 Turbo) can often match or even outperform prompting a larger foundation equivalent (such as GPT-4) whilst being smaller and requiring fewer examples to contextualise the prompt (by up to 90%) and hence cheaper to run.

Of course, fine-tuning and prompt engineering are not mutually exclusive, so there may be some benefit in fine-tuning a model generically for the domain, and then using it to develop solutions for each task via a combination of in-context learning and instruction prompting.

In particular, fine-tuning doesn’t increase domain-level knowledge, so reducing hallucinations might require adopting techniques such as instruction prompting, in-context learning and RAG/embedding, the latter also being beneficial where responses need to be verifiable for legal or regulatory reasons.

Essentially, the choice of approach will very much come down to use case. If the aim is to deliver a natural language search/recommendation capability for use with Enterprise data, a good approach would be to employ semantic embeddings within a RAG framework. Such an approach is highly scalable for dealing with a large database of documents, and able to retrieve more relevant content (via vector search) as well as being more cost-effective than fine-tuning.

Conversely, in the case of a Customer Support chatbot, fine-tuning the model to exhibit the right behaviours and tone of voice will be important, and could then be combined with in-context learning/RAG to ensure the information it has access to is up-to-date.

Choosing a foundational LLM

There are a range of foundational models to choose from with well-known examples coming from OpenAI (GPT-3.5), Google (PaLM 2), Meta (LLama2), Anthropic (Claude 2), Cohere (Command), Databricks (Dolly 2.0), and Israel’s AI21 Labs, plus an increasingly large array of open source variants that have often been fine-tuned towards particular skillsets.

Deployment on-prem provides the Enterprise with more control and privacy, but increasingly a number of players are launching cloud-based solutions that enable Enterprises to fine-tune a model without comprising the privacy of their data (in contrast to the public use of ChatGPT, for example).

OpenAI, for instance, have recently announced availability for fine-tuning on GPT-3.5 Turbo, with GPT-4 coming later this year. For a training file with 100,000 tokens (e.g., 50 examples each with 2000 tokens), the expected cost might be as little as ~$2.40, so experimenting with fine-tuning is certainly within the reach of most Enterprises albeit with the ongoing running costs of using OpenAI’s APIs for utilising the GPT model.

If an Enterprise doesn’t need to fine-tune, OpenAI now offer ChatGPT Enterprise, based on GPT-4, and with an expanded context window (32k tokens), better performance (than the public ChatGPT) and guaranteed security for protecting the Enterprise’s data.

Alternatively, Microsoft have teamed up with Meta to support Llama 2 on Azure and Windows, and for those that prefer more flexibility, Hugging Face has become by far the most popular open source library to train and fine-tune LLMs (and other modalities).

As mentioned previously, players are also bringing to market LLMs pre-trained for use within a particular domain; for example: BloombergGPT for finance; Google’s Med-PaLM-2 for helping clinicians determine medical issues within X-rays and Sec-PaLM which was tweaked for cybersecurity use cases; Salesforce’s XGen-7B family of LLMs for sifting through lengthy documents to extract data insights, or their Einstein GPT (based on ChatGPT) for use with CRM; IBM’s watsonx.aigeospatial foundation model for Earth observation data; AI21 Labs hyper-optimized task-specific models for content management or expert knowledge systems; Harvey AI for generating legal documents etc.

‘Agents’ take the capabilities of LLMs further still by taking a stated goal from the user and combining LLM capabilities with search and other functionality to complete the task — there are a number of open source projects innovating in this area (AutoGPT, AgentGPT, babyagi, JARVIS, HuggingGPT), but also commercial propositions such as Dust.

It’s a busy space… so what are the opportunities (if any) for startups to innovate and claim a slice of the pie?

Uncovering the opportunities

Perhaps not surprisingly given the rapid advancements that have been achieved over the past 12mths, attention in the industry has very much focused on deriving better foundational models and delivering the immense compute resources and data needed to train them, and consequently has created eye-wateringly high barriers for new entrants (Inflection AI recently raising $1.3bn to join the race).

Whilst revenues from offering foundational models and associated services look promising (if you believe the forecasts that OpenAI is set to earn $1bn over the next 12mths), value will also emerge higher up the value stack, building on the shoulders of giants so to speak, and delivering solutions and tools targeted towards specific domains and use cases.

Success at this level will be predicated on domain experience as well as delivering a toolchain or set of SaaS capabilities that enable Enterprises to quickly embrace LLMs, combine them with their data, and generate incremental value and a competitive advantage in their sector.

In stark contrast to the Big Data and AI initiatives in the past that have delivered piecemeal ‘actionable insights’, LLMs have the potential of unlocking comprehensive intelligence, drawing on every documented aspect of a given business, and making it searchable and accessible through natural language by any employee rather than being constrained by the resources of corporate Business Intelligence functions.

But where might startups go hunting for monetisable opportunities?

One potential option is around embeddings — noisy, biased, or poorly-formatted data can lead to suboptimal embeddings resulting in reduced performance, so is a potential micro-area for startups to address: developing a proposition, backed-up with domain-specific experience, and crafting an attractive niche in the value chain helping businesses in targeted sectors.

Another area is around targeted, and potentially personalised, augmentation tools. Whilst the notion of GenAI/LLMs acting as copilots to augment and assist humans is often discussed in relation to software development (GitHub Copilot; StarCoder), it could equally assist workers across a multitude of everyday activities. Language tasks are estimated to account for 62% of office workers’ time, and hence there is in theory huge scope for decomposing these tasks and automating or assisting them using LLM copilots. And just as individuals personalise and customise their productivity tools to best match their individual workflows and sensibilities, the same is likely to apply for LLM copilots.

Many expect that it will turn into an AI gold rush, with those proving commercial value (finding the gold) or delivering the tools to help businesses realise this value (picks & shovels) earning early success and with a chance of selling out to one of the bigger players keen to do a land grab (e.g., Salesforce, Oracle, Microsoft, GCP, AWS etc.) and before the competition catches up.

Defensibility though is likely to be a challenge, at least in the pure sense of protecting IP, and perhaps is reserved for those with access to domain-specific data sets that gives them an edge — Bloomberg, for instance, had the luxury of training their GPT model using their own repository of financial documents spanning forty years (a massive 363 billion tokens).

Takeaways

Foundational LLMs have come a long way, and can be used across a dazzling array of natural language tasks.

And yet when it comes to Enterprise applications, their knowledge is static and therefore out of date, they’re unable to state their source (given the statistical nature of how LLMs produce outputs), and they’re liable to deliver incorrect factual responses (hallucinations).

To do well in an Enterprise setting they need to be provided with detailed and appropriate context, and adequately guided.

Industry and academia are now working diligently to address these needs, and this article has outlined some of the different techniques being developed and employed.

But LLMs are still an immature technology, hence developers and startups that understand the technology in depth are likely to be the ones best able to build more effective applications, and build them faster and more easily, than those who don’t — this, perhaps, is the opportunity for startups.

As stated by OpenAI’s CEO Sam Altman, “Writing a really great prompt for a chatbot persona is an amazingly high-leverage skill and an early example of programming in a little bit of natural language”.

We’re entering the dawn of natural language programming…