- Welcome to the InstructLab Taxonomy
- Learning
- Getting Started with Skill Contributions
- Getting Started with Knowledge Contributions
- Taxonomy tree layout
- Contribute knowledge and skills to the taxonomy!
InstructLab 🐶 uses a novel synthetic data-based alignment tuning method for Large Language Models (LLMs.) The "lab" in InstructLab 🐶 stands for Large-Scale Alignment for ChatBots [1].
The LAB method is driven by taxonomies, which are largely created manually and with care.
This repository contains a taxonomy tree that allows you to create models tuned with your data (enhanced via synthetic data generation) using LAB 🐶 method.
[1] Shivchander Sudalairaj*, Abhishek Bhandwaldar*, Aldo Pareja*, Kai Xu, David D. Cox, Akash Srivastava*. "LAB: Large-Scale Alignment for ChatBots", arXiv preprint arXiv: 2403.01081, 2024. (* denotes equal contributions)
Learn about the concepts of "skills" and "knowledge" in our InstructLab Community Learning Guide.
Skills require a much smaller volume of content than knowledge contributions. An entire skill contribution to the taxonomy tree can be just a few lines of YAML in the qna.yaml
file ("qna" is short for "questions and answers") and an attribution.txt
file for citing sources.
Your skills contribution pull requests must include the following:
- A
qna.yaml
that contains a set of key/value entries with the following keys- Each
qna.yaml
file requires a minimum of five question and answer pairs.
- Each
- An
attribution.txt
that includes the sources for the information used in theqna.yaml
Tip
The skill taxonomy structure is used in several ways:
- To select the right subset of the taxonomy to use for data generation.
- To determine the interpretability by human contributors and maintainers.
- As part of the prompt to the LLM used to generate synthetic samples.
Important
There is a limit to how much content can exist in the question/answer pairs for the model to process. Due to this, only add a maximum
of around 2300 words to your question and answer seed example pairs in the qna.yaml
file.
Taxonomy skill files must be a valid YAML file named qna.yaml
. Each qna.yaml
files contains a set of key/value entries with the following keys:
version
: The value must be the number 2. Requiredtask_description
: A description of the skill. Requiredcreated_by
: The GitHub username of the contributor. Requiredseed_examples
: A collection of key/value entries. New submissions should have at least five entries, although older files may have fewer. Requiredcontext
: Grounded skills require the user to provide context containing information that the model is expected to take into account during processing. This is different from knowledge, where the model is expected to gain facts and background knowledge from the tuning process. The context key is optional for freeform skills.question
: A question for the model. Requiredanswer
: The desired response from the model. Required
Other keys at any level are currently ignored.
To make the qna.yaml
files easier and faster for humans to read, it is recommended to specify version
first, followed by task_description
, then created_by
, and finally seed_examples
.
In seed_examples
, it is recommended to specify context
first (if applicable), followed by question
and answer
.
Example qna.yaml
version: 2
task_description: <string>
created_by: <string>
seed_examples:
- question: <string>
answer: |
<multi-line string>
- context: |
<multi-line string>
question: <string>
answer: |
<multi-line string>
...
Then, you create an attribution.txt
file that includes the sources of your information. These can also be self authored.
Example attribution.txt
[Link to source]
[Link to work]
[License of the work]
[Creator name]
For more information on what to include in your attribution.txt
file, see For your attribution.txt file in CONTRIBUTING.md.
If you have not written YAML before, don't be intimidated - it's just text.
Tip
- Spaces and indentation matter in YAML. Two spaces to indent.
- Don't use tabs!
- Be careful to not have trailing spaces at the end of a line.
- Each example in
seed_examples
begins with a "-". Place this "-" in front of the first field (question
orcontext
). The remaining keys in the example should not have this "-". - Some special characters such as " and ' need to be escaped with backslash. This is why some of the lines for keys in the example YAML start the value with the '|' character followed a new line and then an indented multi-line string. This character disables all of the special characters in the value for the key. You might also want to use the '|' character for multi-line strings.
- Consider quoting all values with " to avoid surprising YAML parser behavior
(e.g. Yes answer can be interpreted by the parser as a boolean of
True
value, unless "Yes" is quoted.) - See https://yaml-multiline.info/ for more info.
It is recommended that you lint, or verify your YAML using a tool. One linter option is yamllint.com. You can copy/paste your YAML into the box and click Go to have it analyze your YAML and make recommendations. Online tools like prettified and yaml-validator can automatically reformat your YAML to adhere to our yamllint
PR checks, such as breaking lines longer than 120 characters.
version: 2
task_description: 'Teach the model how to rhyme.'
created_by: juliadenham
seed_examples:
- question: What are 5 words that rhyme with horn?
answer: warn, torn, born, thorn, and corn.
- question: What are 5 words that rhyme with cat?
answer: bat, gnat, rat, vat, and mat.
- question: What are 5 words that rhyme with poor?
answer: door, shore, core, bore, and tore.
- question: What are 5 words that rhyme with bank?
answer: tank, rank, prank, sank, and drank.
- question: What are 5 words that rhyme with bake?
answer: wake, lake, steak, make, and quake.
Seriously, that's it.
Here is the location of this YAML in the taxonomy tree. Note that the YAML file itself, plus any added directories that contain the file, is the entirety of the skill in terms of a taxonomy contribution:
[...]
└── writing
└── freeform
| └── haikus <=== here it is :)
| | └── qna.yaml
| | attribution.txt
│ ├── debate
│ │ └── qna.yaml
| | attribution.txt
│ ├── legal
│ │ ├── agreement
│ │ | └── qna.yaml
| | | attribution.txt
[...]
Remember that grounded compositional skills require additional context and include a context
field.
This example snippet assumes the GitHub username mairin
and shows some of the question/answer pairs present in the actual file:
version: 2
task_description: |
This skill provides the ability to read a markdown-formatted table.
created_by: mairin # Use your GitHub username; only one creator supported
seed_examples:
- context: |
| **Breed** | **Size** | **Barking** | **Energy** |
|----------------|--------------|-------------|------------|
| Afghan Hound | 25-27 in | 3/5 | 4/5 |
| Labrador | 22.5-24.5 in | 3/5 | 5/5 |
| Cocker Spaniel | 14.5-15.5 in | 3/5 | 4/5 |
| Poodle (Toy) | <= 10 in | 4/5 | 4/5 |
question: |
Which breed has the most energy?
answer: |
The breed with the most energy is the Labrador.
- context: |
| **Name** | **Date** | **Color** | **Letter** | **Number** |
|----------|----------|-----------|------------|------------|
| George | Mar 5 | Green | A | 1 |
| Gráinne | Dec 31 | Red | B | 2 |
| Abigail | Jan 17 | Yellow | C | 3 |
| Bhavna | Apr 29 | Purple | D | 4 |
| Rémy | Sep 9 | Blue | E | 5 |
question: |
What is Gráinne's letter and what is her color?
answer: |
Gráinne's letter is B and her color is red.
- context: |
| Banana | Apple | Blueberry | Strawberry |
|--------|------------|-----------|------------|
| Yellow | Red, Green | Blue | Red |
| Large | Medium | Small | Small |
| Peel | Peel | No peel | No peel |
question: |
Which fruit is blue, small, and has no peel?
answer: |
The blueberry is blue, small, and has no peel.
[...]
└── extraction
└── inference
| └── qualitative
| | ├── sentiment
| | | └── qna.yaml
| | | attribution.txt
| | └── tone_and_style
| | └── qna.yaml
| | attribution.txt
│ ├── quantitative
│ │ ├── table_analysis <=== here it is :)
│ | | └── qna.yaml
│ │ │ attribution.txt
│ │ ├── word_frequency
│ │ │ └── qna.yaml
│ │ │ attribution.txt
[...]
Important
Upon release, the taxonomy repository is only accepting contributions from Wikipedia and is capped at 50 contributions. If you want to add knowledge to the taxonomy repository, please fill out this InstructLab Knowledge Submission Registration form and await acceptance! Please do not add contributions if you do not receive the confirmation email. Thank you!
While skills are foundational or performative, knowledge is based more on answering questions that involve facts, data, or references.
Knowledge in the taxonomy tree consists of a few more elements than skills:
- Each knowledge node in the tree has a
qna.yaml
, similar to the format of theqna.yaml
for skills. - ⭐ Knowledge submissions require you to create a Git repository, can be with GitHub, that contains the markdown files of your knowledge contributions. These contributions in your repository must use the markdown (.md) format.
- The
qna.yaml
includes parameters that contain information from your repository.
Tip
Guidelines for Knowledge contributions
- Submit the most up-to-date version of the document
- All submissions must be text, images will be ignored
- Do not use tables in your markdown freeform contribution
Important
There is a limit to how much content can exist in the question/answer pairs for the model to process. Due to this, only add a maximum
of around 2300 words to your question and answer seed example pairs in the qna.yaml
file.
Each qna.yaml
file requires a minimum of five question-answer pairs. The qna.yaml
format must include the following fields:
version
: The value must be the number 2.task_description
: An optional description of the knowledge.created_by
: Your GitHub username.domain
: Category of the knowledge.seed_examples
: Five or more examples sourced from the provided knowledge documents.question
: A question for the model. This key is required.answer
: The desired response from the model. This key is required.
document
: The source of your knowledge contribution.repo
: The URL to your repository that holds your knowledge markdown files.commit
: The SHA of the commit in your repository with your knowledge markdown files.patterns
: A list of glob patterns specifying the markdown files in your repository. Any glob pattern that starts with*
, such as*.md
, must be quoted due to YAML rules. For example,"*.md"
.
version: 2
task_description: 'Teach the model the results of the 2024 Oscars'
created_by: juliadenham
domain: pop_culture
seed_examples:
- question: When did the 2024 Oscars happen?
answer: |
The 2024 Oscars were held on March 10, 2024.
- question: What film had the most Oscar nominations in 2024?
answer: |
Oppenheimer had 13 Oscar nominations.
- question: Who presented the 2024 Oscar for Best Original Screenplay and Best Adapted Screenplay?
answer: |
Octavia Spencer presented the award for Best Original Screenplay and Best Adapted Screenplay at the 2024 Oscars.
- question: Who hosted the 2024 Oscars?
answer: |
Jimmy Kimmel hosted the 96th Academy Awards ceremony.
- question: At the 2024 Oscars, who were the nominees for best director and who won?
answer: |
The nominees for director at the 2024 Oscars was Christopher Nolan for Oppenheimer,
Justine Triet for Anatomy of a Fall, Martin Scorsese for Killers of the Flower Moon,
Yorgos Lanthimos for Poor Things, and Jonathan Glazer for The Zone of Interest.
Christopher Nolan won best director for Oppenheimer.
- question: Did Billie Eilish perform at the 2024 Oscars?
answer: |
Yes Billie Eilish performed "What Was I Made For?" from Barbie at the 2024 Oscars.
document:
repo: https://github.com/juliadenham/oscars2024_knowledge.git
commit: e1744af
patterns:
- oscars2024_results.md
Example attribution.txt
file
Title of work: 96th Academy Awards
Link to work: https://en.wikipedia.org/wiki/96th_Academy_Awards
License of the work: CC-BY-SA-4.0
Creator names: Wikipedia Authors
This knowledge example references one markdown file: oscars2024_results.md
. You can also add multiple files for knowledge contributions.
Note
Due to the higher volume, it will naturally take longer to receive acceptance for a knowledge contribution pull request than for a skill pull request. Smaller pull requests are simpler and require less time and effort to review.
What might these markdown files look like? They can be freeform. Here's what a
snippet of oscars2024_results.md
might look like in your Git repository.
# 96th Academy awards
The **96th Academy Awards** ceremony, which was presented by the
[Academy of Motion Picture Arts and
Sciences](Academy_of_Motion_Picture_Arts_and_Sciences "wikilink")
(AMPAS), took place on March 10, 2024, at the [Dolby
Theatre](Dolby_Theatre "wikilink") in
[Hollywood](Hollywood,_Los_Angeles "wikilink"), Los Angeles.[1] During
the gala, the AMPAS presented [Academy
Awards](Academy_Awards "wikilink") (commonly referred to as Oscars) in
23 categories honoring [films released in
2023](2023_in_film "wikilink"). Comedian [Jimmy
Kimmel](Jimmy_Kimmel "wikilink") hosted the show for the fourth time.
The nominations were announced on January 23, 2024.
*[Oppenheimer](Oppenheimer_(film) "wikilink")* led with 13 nominations,
followed by *[Poor Things](Poor_Things_(film) "wikilink")* and *[Killers
of the Flower Moon](Killers_of_the_Flower_Moon_(film) "wikilink")* with
11 and 10, respectively.[2][3][4] *Oppenheimer* won a leading seven
awards, including [Best
Picture](Academy_Award_for_Best_Picture "wikilink") and [Best
Director](Academy_Award_for_Best_Director "wikilink")
[..]
In the taxonomy repository, here's what the previously referenced knowledge might look like in the tree:
[...]
└── knowledge
└── textbooks
├── culture
│ └── movies
│ └── awards
│ ├── oscars <=== here it is :)
│ │ └── qna.yaml
| | attribution.txt
│ └── golden_globes_movies
│ └── qna.yaml
| attribution.txt
[...]
For more information on what to include in your attribution.txt
file, see For your attribution.txt file in CONTRIBUTING.md.
You can organize the knowledge markdown files in your repository however you want. You just need to ensure the YAML is pointing to the correct file.
The taxonomy tree is organized in a cascading directory structure. At the end of each branch, there is a YAML file (qna.yaml) that contains the examples for that domain. Maintainers can decide to change the names of the existing branches or to add new branches.
Important
Folder names do not have spaces.
Below is an illustrative directory structure to show this layout:
.
└── writing
├── freeform
│ ├── brainstorming
│ │ ├── idea_generation
│ │ │ └── qna.yaml
│ │ │ attribution.txt
│ │ ├── refute_claim
│ │ │ └── qna.yaml
│ │ │ attribution.txt
│ ├── prose
│ │ ├── articles
│ │ │ └── qna.yaml
│ │ │ attribution.txt
│ │ ├── emails
│ │ │ ├── formal
│ │ │ │ └── qna.yaml
│ │ │ │ attribution.txt
│ │ │ └── informal
│ │ │ └── qna.yaml
│ │ │ attribution.txt
└── grounded
├── editing
│ ├── grammar
│ │ └── qna.yaml
│ │ attribution.txt
│ └── spelling
│ └── qna.yaml
│ attribution.txt
└── summarization
└── wiki_insights
└── concise
└── qna.yaml
attribution.txt
For an extensive example of this layout see, taxonomy_tree_layout in the documentation folder.
The ability to contribute to a Large Language Model (LLM) has been difficult in no small part because it is difficult to get access to the necessary compute infrastructure.
This taxonomy repository will be used as the seed to synthesize the training data for InstructLab-trained models. We intend to retrain the model(s) using the main branch following InstructLab's progressive training on a regular basis. This enables fast iteration of the model(s), for the benefit of the open source community.
By contributing your skills and knowledge to this repository, you will see your changes built into an LLM within days of your contribution rather than months or years! If you are working with a model and notice its knowledge or ability lacking, you can correct it by contributing knowledge or skills and check if it's improved after your changes are built.
While public contributions are welcome to help drive community progress, you can also fork this repository under the Apache License, Version 2.0, add your own internal skills, and train your own models internally. However, you might need your own access to significant compute infrastructure to perform sufficient retraining.
You can contribute to the taxonomy in the following two ways:
- Adding new examples to existing leaf nodes:
- Adding new branches/skills corresponding to the existing domain:
For more information, see the Ways of contributing to the taxonomy repository documentation.
To contribute to this repo, you'll use the Fork and Pull model common in many open source repositories. You can add your skills and knowledge to the taxonomy in multiple ways; for additional information on how to make a contribution, see the Documentation on contributing. You can also use the following guides to help with contributing:
- Contributing using the GitHub webpage UI.
- Contributing knowledge to the taxonomy in the Knowledge contribution guidelines.
This taxonomy repository will be used as the seed to synthesize the training data for InstructLab-trained models. We intend to retrain the model(s) using the main branch as often as possible (at least weekly). Fast iteration of the model(s) benefits the open source community and enables model developers who do not have access to the necessary compute infrastructure.