/llmi

Large-Language-Model to Machine Interface project.

GNU General Public License v3.0GPL-3.0

llmi

Large-Language-Model to Machine Interface Open-Source project.

Goal

Large Language Models (LLMs) are very good at some things (writing, summarizing, answering accurate questions), but also remarkably bad at others (real-time information, complex logic, information formatting, tool usage, verification, structured/long output).

Some of those things they are bad at, they could get better at if they were able to get help from tools/machines that currently

This project aims to create:

  • Specifications
  • Training datasets
  • Large language models
  • And tools

in order to create LLMs with the ability to use computers to improve their abilities.

Here are some of the abilities this project aims to implement:

  • LLM as a library/function-call.
  • TODO

Implementation

The plan here is, for each/all of the features, to:

  1. Write down a specification of what the feature should be able to accomplish
  2. Plan and document how to accomplish/implement it.
  3. Create a training dataset for this feature.
  4. Once all datasets exist, train a LLM using these datasets.
  5. Test and improve the LLM.
  6. Write tools that are able to take advantage of the new features.

Principles

  1. Transparency/fallback: The content should still look reasonably readable to a human if the tags are removed and no processing is done on them.
  2. Some of these features can probably be implemented (though maybe not very well) using existing models. We should try that wherever possible, as it could be helpful for writing the prompts/datasets.

Features

This is a presentation of each feature this project aims to implement.

PRs for extra features (and any contribution) are extremely welcome:

Recursive redaction.

Goal: Allow LLMs to write large bodies of text recursively, enabling chapters, sub-chapters etc to be writtens separately (either in series or in parralel) and then joined together.

This would allow writing much longer content than is currently allowed by LLMs, which are limited by their context window, the only disadvantage being this would only be possible for "structured" content.

The way this would be implemented is, the LLM would first generate a table of contents for the body of text, and then in each chapter, insert a tag asking the system to write that chapter.

For example, if asked:

write a dissertation explaining why the earth is not flat

The LLM would generate something in this format:

    # 1. Introduction

      <llmi:chapter context=true title="Introduction">
        The debate surrounding the shape of the Earth seems like an archaic one, but it has resurfaced in modern times due to social media and the availability of misleading information online.
        It's crucial to address these misconceptions by leveraging empirical evidence and scientific reasoning.
      </llmi:chapter>

    # 2. History of Flat Earth Theory

      <llmi:chapter context=true title="History of Flat Earth Theory">
        The flat Earth theory has its roots in ancient civilizations but was largely debunked by the advent of the Scientific Revolution.
        Philosophers and scientists like Pythagoras, Aristotle, and Copernicus presented evidence supporting the Earth's spherical shape, which has stood the test of time.
      </llmi:chapter>

    # 3. Fundamental Laws of Physics

      <llmi:chapter context=true title="Fundamental Laws of Physics">

    [etc...]

Then, as the output is being generated by running the LLM, each time a <llmi:chapter> tag is generated, when the </llmi:chapter> tag is found, the system will:

  1. Create a new prompt from that tag
  2. Use it to generate the new chapter
  3. Replace, in the output, the <llmi:chapter> tag with the generated output.

This can be done either in series (we wait for the generation to finish to move on to the rest of the content), or in parralel (each time we hit a tag, we start a new process but continue going on the main content).

For example, here for the second <llmi:chapter> tag, when it is hit, the following prompt would be generated:

An assistant was requested to write a text, structured by chapter, based on the prompt:

« write a dissertation explaining why the earth is not flat »

It decided one of the chapters would be named:

« History of Flat Earth Theory »

It wrote a short summary of the content of this chapter:

«  The flat Earth theory has its roots in ancient civilizations but was largely debunked by the advent of the Scientific Revolution.
   Philosophers and scientists like Pythagoras, Aristotle, and Copernicus presented evidence supporting the Earth's spherical shape, which has stood the test of time.
»

Please write a long version of this chapter.

If necessary, use <llmi:chapter> tags to recursively go down the table of contents structure

The system would ask the LLM to generate the new chapter according to this new prompt.

Once the chapter written, the <llmi:chapter> tag would be replaced by the chapter itself in the output.

This means once the entire process is finished, the output would be an entire dissertation, with each chapter complete, even if the content is much longer than the context/attention window of the LLM.

Note: This is not limited to documentation-like or dissertation-like structures, you could create very large structures such as "a tree of all education with «this is needed to learn that» relationships".

Thinking with code.

Goal: Allow the LLM to solve mathematics or logic problems by writing and executing code.

(edit: I guess this is coming to MS Copilot: https://www.theverge.com/2023/12/5/23989052/microsoft-copilot-gpt-4-turbo-openai-models-code-interpreter-feature so it mustn't be a terrible idea... )

LLMs can struggle with solving math or logic problems by themselves, however they are known to be able to write code.

If we allowed them to transform these problems into code (after training them to learn to do so), and then access the result of executing that code, it would enable them to have a higher fidelity in solving these problems.

For example, if provided with the following prompt:

You have 100 fruits in total, some of the fruits are oranges, and some of the fruits are apples.
You have 60 more apples than you have oranges.
Provide the total number of each fruit.

The output from the LLM would be something like:

Rewriting the problem as code, would give us something like this:

<llmi:code id=1 execute=true>
# Initialize a variable for the total number of fruits
total_fruits = 100

# According to the problem, the number of apples is 60 more than the number of oranges.
# Let's assume the number of oranges is x
# Then the number of apples would be x + 60

# Equation to solve: x + (x + 60) = 100

# Solve for the number of oranges (x)
number_of_oranges = (total_fruits - 60) / 2

# Calculate the number of apples using the equation: number of apples = x + 60
number_of_apples = number_of_oranges + 60

# Print the results
print(f"Number of apples: {number_of_apples}")
print(f"Number of oranges: {number_of_oranges}")
</llmi:code>

This code would generate the following output:

<llmi:output id=1> </llmi:output>

As the LLM would generate this output, as it hits/sees the </llmi:code> tag, it would execute everything in the <llmi:code> tag, and store the output.

Then, when it hits the <llmi:output> tag, it would get the output from the <llmi:code> tag, and insert the output inside the tag, so that:

<llmi:output id=1> </llmi:output>

Woud become:

<llmi:output id=1>
Number of apples: 80
Number of oranges: 20
</llmi:output>

This new content for the tag would then become a part of the current context for the generation (ie it would be fed into the LLM for the generation of further tokens).

This means right after this tag, the LLM would now be able to continue the generation with access to the output of the code, by writing something like:

Thus, the number of oranges is 20, and the number of apples is 80, which is indeed 60 more.

Through this process, the LLM has been able to "use" the execution of code to solve the problem.

Another example, spreadsheet engine.

Let's say I want a list of pairs of acid/base, the reaction of which produces both carbon dioxide and water.

While this can be a difficult task for LLMs today, it is much easier for them to write code, SQL requests or create a spreadsheet that solves the problem.

So, we could ask the LLM:

Give me a list of pairs of acid/base, the reaction of which produces both carbon dioxide and water.

And it could be trained to reply something like:


Here is a table of various combinations of acids and bases, and the resulting products of their reactions:

<llmi:database table="reactions">
  <tr>
    <th>Acid</th>
    <th>Base</th>
    <th>Products</th>
    <th>Produces CO2</th>
    <th>Produces H2O</th>
  </tr>
  <tr>
    <td>Hydrochloric acid</td>
    <td>Sodium bicarbonate</td>
    <td>Sodium chloride, CO2, water</td>
    <td>Yes</td>
    <td>Yes</td>
  </tr>
  <tr>
    <td>Sufuric Acid</td>
    <td>Potassium Hydroxide</td>
    <td>Potassium sulfate, water</td>
    <td>No</td>
    <td>Yes</td>
  </tr> 
  [etc...]
</llmi:database>

Among these, here is a list of those that produce both carbon dioxide and water:

<llmi:request execute=true>
    SELECT * from `reactions` where `Produces CO2` = 'Yes' and `Produces H2O` = 'Yes'
</llmi:request>

...

As the </llmi:request> tag is hit, the system would execute the SQL request, and insert the result in the output, then continue generating from there, with the LLM now "aware" of the result, and able to comment on it.

Note: Here it used a SQL request, as it's pretty adequate to the question/manipulating tabular data, but it could potentially have used code or a spreadsheet instead.

TODO: From the «GPT4 can't reason» paper.

TODO: You are given the following premises: Someone who lives in Dreadbury Mansion killed Aunt Agatha. The only people who live in Dreadbury Mansion are Aunt Agatha, the butler, and Charles. A killer always hates his victims, and is never richer than his victims. Charles hates no one that Aunt Agatha hates. Aunt Agatha hates everyone except the butler. The butler hates everyone not richer than Aunt Agatha. The butler hates everyone Aunt Agatha hates. No one hates everyone. Aunt Agatha is not the butler. On the basis of this information, determine who killed Aunt Agatha and give a detailed proof that your conclusion follows from the premise.

Machine readable content.

This one is more about what the LLM should not do, than about what it should do.

Goal: When asked the LLM should generate output in a predictible and repeatable format that is easy to interprate for a machine.

This is actually a feature that is going to be required for some of the other features to work properly.

For example, if the LLM gets the prompt:

Give the tone of the following comment: « Lol, sure, **that happened** ».

A LLM would typically reply something like:

Certainly! The tone of the comment "Lol, sure, that happened" appears to be sarcastic. The phrase "that happened" is often used online to express disbelief or skepticism, implying that the speaker does not actually believe the event in question took place. Adding "Lol" and "sure" further accentuates the sarcastic tone.

But if you are writing a program that is meant to process/categorize the tone of millions of comments (or one that is meant to write chapters recursively, see above), this is not the output you want, this is nearly impossible to parse. It also wastes ressources by repeating the original comment etc. We only want one word here.

Instead, a much better output would be, if we ask:

Give the tone of the following comment: « Lol, sure, **that happened** », in a machine-parsable format.

The output would then be:

<llmi:output>
    sarcastic
    <llmi:reason> The phrase "that happened" is often used online to express disbelief or skepticism, implying that the speaker does not actually believe the event in question took place. Adding "Lol" and "sure" further accentuates the sarcastic tone. </llmi:reason>
</llmi:output>

This is much easier to parse. You could also limit the number of possible replies by providing a list of possible values, etc...

Alternatively, a version that provides both machine-parsable and user-friendly output is possible, for example:

Certainly!
The tone of the comment <llmi:original>"Lol, sure, that happened"</llmi:original> appears to be <llmi:output> sarcastic </llmi:output>.
<llmi:explanation> The phrase "that happened" is often used online to express disbelief or skepticism, implying that the speaker does not actually believe the event in question took place. Adding "Lol" and "sure" further accentuates the sarcastic tone. </llmi:explanation>

By hidding the tags, the output is perfectly presentable as a reply to an end-user, and by using a DOM parser to get to the <llmi:output> tag, we have perfectly fine machine-parsable content.

Getting extra context

Goal: Enable the system to grab extra content/information, and provide it to the context of the LLM itself without this extra information being visiable to the end user.

This is another "replacement" type of tag, it's pretty simple.

Say you ask the model:

Tell me about Ethylenetetracarboxylic acid

The output from another LLM would normally then something like:

As of my last update in January 2022, Ethylenetetracarboxylic acid doesn't appear to be a standard or widely-recognized chemical compound in the scientific literature. It seems like the name suggests an organic acid that would contain an ethylene group and four carboxylic acid groups, but such a compound doesn't seem to be commonly referenced or studied.

In organic chemistry, names of compounds are usually systematic and based on their molecular structure according to rules set by the International Union of Pure and Applied Chemistry (IUPAC). Ethylenetetracarboxylic acid would imply a two-carbon ethylene
[etc...]

However, in the currently proposed system, instead, the output from the LLM would be:

<llmi:wikipedia page="Ethylenetetracarboxylic_acid"></llmi:wikipedia>

(note, here Wikipedia is used as an example, but other sources could be used, and this could be significantly improved, for example adding information to the context until its maximum length is reached, which would enable the model/system to reach it's "maximum ability to get extra info").

Now, before the LLM keeps generating the tokens after this tag, the system would "interpret" the tag, by filling it with the content of the Wikipedia page.

This means the context of the LLM would , as soon as it hits </llmi:wikipedia>, become:

<llmi:wikipedia page="Ethylenetetracarboxylic_acid">
Ethylenetetracarboxylic acid is an organic compound with formula C6H4O8, or (HO(OC)-)2C=C(-(CO)OH)2.
By removal of four protons, the acid yields the anion C6O4−8, ethylenetetracarboxylate, which is one of the oxocarbon anions (consisting solely of oxygen and carbon). By loss of 1 through 3 protons, it forms the anions C6H3O−8, C6H2O2−8, and C6HO3−8, called respectively trihydrogen-, dihydrogen-, and hydrogenethylenetetracarboxylate. The same names are used for the corresponding esters.
The acid can be obtained by hydrolysis of tetraethyl ethylenetetracarboxylate, which in turn can be obtained from diethyl dibromomalonate with sodium iodide.
Ethylenetetracarboxylic dianhydride, a twofold acid anhydride of this compound, can be formed by direct dehydration at high temperature.
</llmi:wikipedia>

(note, in the final output of the system to the end user, the <llmi:wikipedia> tag would be hidden.)

From there, with this new context that contains new/extra information, the LLM would be able to continue generating the tokens after <llmi:wikipedia> in a more informed manner:

Ethylenetetracarboxylic acid is a known organic compound with the chemical formula C6H4O8C6​H4​O8​ or (HO(OC)−)2C=C(−(CO)OH)2(HO(OC)−)2​C=C(−(CO)OH)2​. This compound can yield various anions by the removal of one to four protons, classifying it among the oxocarbon anions—ions consisting solely of carbon and oxygen.

You can obtain Ethylenetetracarboxylic acid through the hydrolysis of its ester, tetraethyl ethylenetetracarboxylate. This precursor can be synthesized from diethyl dibromomalonate using sodium iodide. Additionally, Ethylenetetracarboxylic dianhydride, a twofold acid anhydride of this compound, can be formed by dire

As you can see, without the Wikipedia page, the model (this was tested with a real model) was unaware of the compound, and thought it was "not standard", but once the Wikipedia page was provided in full, it provided a more up-to-date/complete answer.

This "live" replacement of tags in token output is fairly simple and cheap, and the only difficult part here is to "teach" the model, through new training data, to learn to add these new tags, so it can benefit from them.

Checking for accuracy.

Goal: Enable the LLM to answer questions of accurateity in a machine-parsable format, using extra context if needed.

For example, asking a normal LLM to fact-check something looks like this:

Is the statement « 3-Hydroxy-2-naphthoic_acid a red crystaline powder at room temperature » correct?

We'd get something like:

As of my last update in January 2022, I don't have specific information to confirm or refute the statement that 3-Hydroxy-2-naphthoic acid is a red crystalline powder at room temperature. For the most accurate and specific information, it's advisable to consult the material safety data sheet (MSDS) or other scientific literature related to this particular compound. These documents usually provide details on the appearance, stability, and other physical and chemical properties of the compound.

However, if we asked our system with the following prompt:

Is the statement « 3-Hydroxy-2-naphthoic_acid a red crystaline powder at room temperature » correct?
Provide the answer in a machine-parsable format.

It would be able to gather information from Wikipedia (see Getting extra context section above):

And because Wikipedia is aware that 3-Hydroxy-2-naphthoic acid is a yellow solid, after the <llmi:wikipedia> tag is replaced with the content of the wikipedia page, the LLM woud be able to reply with:

<llmi:fact accurate=false statement="3-Hydroxy-2-naphthoic_acid a red crystaline powder at room temperature" correction="3-Hydroxy-2-naphthoic_acid a yellow solid at room temperature">
    This is incorrect, 3-Hydroxy-2-naphthoic acid is a yellow solid at room temperature. For the most specific and detailed information, it's advisable to consult the material safety data sheet (MSDS) or other scientific literature related to this particular compound. These documents usually provide comprehensive details on the appearance, stability, and other physical and chemical properties of the compound.
</llmi:fact>

This means the system would be able to fact-check factoids, and provide them in a machine-parsable format. This then enables the next feature in the list, « Integrating fact-checking into normal output ».

Integrating fact-checking into normal output / limiting hallucinations.

Goal: Transparently enable the accurateity of statements in the output of the LLM to be checked, and corrected if needed.

Let's say you ask a LLM:

Tell me about acetic acid

It might reply something like:

Acetic acid, also known as ethanoic acid, is a colorless liquid organic compound with the chemical formula C2H4O2, characterized by a pungent, vinegar-like odor and is the main component of vinegar, making up about 4–8% of its volume,

(this might be accompanied with someting like «fact check each fact in your reply», either in the prompt itself, or the system prompt).

Now with our system, with fact-checking enabled, the output would instead be something like:

Acetic acid,
<llmi:fact statement="Acetic acid is also known as ethanoic acid">also known as ethanoic acid,</llmi:fact>
<llmi:fact statement="Acetic acid is a blue liquid organic compound">is a blue liquid organic compound</llmi:fact>
<llmi:fact statement="Acetic acid has for chemical formula C2H402">with the chemical formula C2H4O2,</llmi:fact>
<llmi:fact statement="Acetic acid has a pangent, vinegar-like odor">characterized by a pungent, vinegar-like odor</llmi:fact>
and is the main component of vinegar, making up about 4–8% of its volume

(there's intentionally an error here, it's an incolor liquid not a blue liquid).

For each of the <llmi:fact> tags, the system would fact-check it, either in parralel or in series, using the LLM itself.

For example, for

<llmi:fact statement="Acetic acid is a blue liquid organic compound">is a blue liquid organic compound</llmi:fact>

We would use the fact-checking feature described in this document, asking:

Is the statement « Acetic acid is a blue liquid organic compound » correct?
Provide the answer in a machine-parsable format.

And get the following output (possibly from Wikipedia or some other source of presumed improved accurateity):

<llmi:fact accurate=false statement="Acetic acid is a blue liquid organic compound" correction="Acetic acid is a colorless liquid organic compound">
    This is false, acetic acid is a colorless liquid.
</llmi:fact>

Now that we know that one of the facts in the output of the LLM is incorrect, we ask the LLM to rewrite the reply taking this information into account, with the following prompt:

The following reply:

«
Acetic acid, also known as ethanoic acid, is a blue liquid organic compound with the chemical formula C2H4O2, characterized by a pungent, vinegar-like odor and is the main component of vinegar, making up about 4–8% of its volume,
»

contains a accurate error, it states:

« Acetic acid is a blue liquid organic compound »

However, that is incorrect, instead:

« Acetic acid is a colorless liquid. »

Please rewrite the reply, taking this corrected information into account.

Which would result in the output:

Acetic acid, also known as ethanoic acid, is an incolor liquid organic compound with the chemical formula C2H4O2, characterized by a pungent, vinegar-like odor and is the main component of vinegar, making up about 4–8% of its volume,

The system identified accurate statements in the reply, fact-checked each of them, and when an error was found, the output was corrected.

The process presented here is a bit over-simplified. Ideally, the fact-checking would happen "on the fly", and we'd change the information in the context as the output is being generated, instead of doing a second "full" call to correct the information.

Hidden reasonning steps.

Augmented output / Callouts

A frequent feature in documentation, "notes" are small apparts in the text noting something that's not directly the content/reply itself, but that is important to know about anyway. In websites they are often displayed inside a box with an icon or specific colors, sometimes called "callouts" ( https://getbootstrap.com/docs/5.0/customize/components/#creating-your-own ).

For example, if you ask a LLM to list the logical fallacies in a given text, it might, at some point, insert into the reply:

It's worth noting that logical fallacies can appear in speeches and writings by individuals across the political spectrum. The presence of a logical fallacy does not necessarily negate the overall point that a speaker is making, but it is a point of interest for those studying rhetoric, logic, or debate.

The system could be trained to recognize these notes and tagging them, enabling:

  1. Displaying them in a special way in HTML/GUI output
  2. Removing them more easily if needed.

Write your own adventure: How to store context information and retrieve it as needed for writing.

The output would become something like:

<llmi:note type="caution" icon="caution-sign">
It's worth noting that logical fallacies can appear in speeches and writings by individuals across the political spectrum. The presence of a logical fallacy does not necessarily negate the overall point that a speaker is making, but it is a point of interest for those studying rhetoric, logic, or debate.
</llmi:note>

Context determination/search.

When you ask a LLM a question, you'd ideally want the LLM to be "aware" of all questions you asked previously.

Some systems, like ChatGPT, implement this by having a "context window": along with each new prompt, they pass along (all/a lot of) previous prompts.

The problem is, this is limited by the context window size for the model, which for Open-Source models, tends to be fairly short.

One thing you could do is, when you get a new prompt X, for each of the previous prompt Yn, you ask the model « Does prompt Yn contain information relevant to answering new prompt X? Write your answer in a form that would be helpful to answer X if passed along with it to the LLM ».

Doing this for all previous prompt is expensive computationally, but :

  1. It can be made parralel (and the cost of executing LLMs will eventually go down, efficiency should go up too)
  2. This is not a magical replacement for sliding context windows, but it can be an extra option for those situations where it is easier to do massive-parralel execution than it is to make the context window for the model larger.

Bringing in large coding projects.

Writing large coding projects as a story/instructions/blog post.

File munipulation instructions.

Spreadsheet manipulation instructions.

Recursively exploring all pro/con arguments of an issue, down to the bottom.

Models working inside docker containers.

Goal: We ask a model to install wordpress. The model glabs the wordpress documentation off the internet, gets it/selected parts of it into context, and then writes the instructions to install wordpress inside a docker container. As the instructions are created, tags to actually run the commands in a docker container are written down in the output, and* interpreted/executed by the system. The final result is a link (localhost:3123 style) that lets the user access the local wordpress installation.

Web interface manipulation (pupeteer), login, subscriptions, etc.

«create an account at examplewebsite.com»

Either by having it write js/ts/pupeteer code directly, or creating a set of dedicated tags for DOM/page manipulation, and using that with a custom model.

HTML for the page as part of the context. Could train a model to "clean up" pages, removing stuff we know for sure won't be useful, so we can feed smaller sizes into the context/prompt.

Additionally, multimodal (image analysis) processing with something like llava (https://www.reddit.com/r/LocalLLaMA/comments/17e855d/llamacpp_server_now_supports_multimodal/) could be used to "look at" the page and provide extra information to the model.

Ambiguous Google searches.

Searching for "ways to filter off activated charcoal particles out of water" gives you info about activated charcoal filters, but not about the actual requested task. A LLM could be taught to create better searches based on the exact request/task, and taught to filter off bad results that are irrelevant to the task.

Picture/link analysis

https://www.reddit.com/r/LocalLLaMA/comments/1750mxn/llamacpp_update_gguf_llava_v15_support_soon/

Teaching models to accept quotes/markup in prompts.

Teaching models to output nice formatting (math, chemical formulas, etc).

Also, fallback to text when copy/pasted, like that

Using specialized models for specific sub-tasks: expert models and expert dispatch.

Ex: code-llama specifically when generating code.

Filter model: Pass-through to detect if format (tags not closed?) is broken, or detect this prog, but then use a model to fix the issue.

Implementation

Note: Mixture of experts/models.

Notes.

Useful links: