Instruction tuning data
Prompts and model-quality answers written by people who know the domain, for supervised fine-tuning by your team.
For assistants that follow instructions
LLM data for AI teams
LLM training data, written and judged by people
LLM training data is the prompts, answers and human judgments that teach a large language model how to behave. We put it together with people from our 30,000+ network, in the 50 most widely spoken languages and deep across Indian ones. Your team does the training.
prompt · written by a person
user Explain a bank overdraft to someone opening their first account.
prompttwo answers · which is better?
1 · preferred Plain, short, gives one example.
2 Accurate, but full of jargon.
checked before delivery
follows guideline second review English and Hindi reason written
reviewed✓ Judged by people reason written for each pick
50 50 languages plus deep Indian coverage
The short answer
What is LLM training data?
LLM training data is the text a large language model learns from and gets checked against: prompts paired with good answers, comparisons that say which answer people prefer, multi-turn conversations and held-back test sets. For a model that has to follow instructions and sound right to real users, most of that data is written or judged by people.
Pretraining uses huge amounts of general text. The data that shapes how a model behaves afterwards is much smaller, and much more deliberate. Someone has to write the good answer, or read two answers and say which one is better, and do it the same way a few thousand times.
We don't train or fine-tune your model. We put the data together with people from our network and check it, and your team does the training.
Prompt, answer and preference labels need the same discipline as any annotation task, and the guide to annotation quality covers what that discipline looks like.
LLM data at IndecalAI
- Kinds
- Prompts and responses, preference comparisons, conversations and evaluation sets.
- Who writes it
- People from a network of 30,000+ people worldwide, matched to the domain.
- Languages
- The 50 most widely spoken languages, with deep coverage of Indian languages.
- Quality
- Written guidelines, a second review and disagreements fed back into the guideline.
- Training
- Done by your team. We provide the data and never train a model.
- Your data
- NDA first, used only for the agreed scope, returned or deleted when you ask.
- Contact
- One named contact who replies within one business day.
- First step
- A small free sample before you commit.
- Company
- Based in Ahmedabad, India, since 2021, with 150+ projects delivered.
AI showcase
Take apart one piece of LLM data
The same kind of prompt, shown as four kinds of LLM data. Tap through them. It's a made-up example, and your guideline decides what the real thing looks like.
Instruction tuning data
A prompt and one good answer, written by a person who knows the subject. The model learns what a helpful reply looks like by seeing thousands of them.
- Written: By people matched to the domain, from banking to healthcare.
- Style: Length, tone and format set in your guideline.
- Checked: A second person reviews a share of every batch.
Preference data for RLHF
People read two or more answers to the same prompt and pick the better one, with a short reason. Your team uses these judgments for RLHF or a similar method.
- Format: Pick the better one, a full ranking or a score, as you need.
- Reasons: A written line on why, when you want it.
- Agreement: Overlap between raters shows where the guideline is unclear.
LLM evaluation data
People rate your model's answers against the targets you set, and results get split by language and topic so weak spots show. See data and model evaluation.
- Scale: Rated on the criteria in your rubric.
- Split: Results by language, topic or user group.
- Held back: Test prompts kept apart from anything used in training.
Multilingual LLM data
Prompts and answers written by native speakers, not run through translation and left there. It covers the 50 most widely spoken languages, and goes deep on Indian ones.
- Written: Natively, or translated and then checked by a native speaker.
- Mixed: Hindi and English in one message, the way people really type.
- Checked: By people who speak the language.
What we put together
The kinds of LLM data we provide
Six kinds of data, all written or judged by people and reviewed before delivery. Most projects mix two or three.
Preference data for RLHF
Two or more answers compared or put in order by people, with a reason, for RLHF and similar methods your team runs.
For answers that match what users prefer
Conversation data
Multi-turn dialogues written or role-played by people, so a model learns to follow a thread. More on conversational AI data.
For chat assistants and support bots
LLM evaluation data
Test prompts and human ratings of your model's answers, split so weak spots show. See data and model evaluation.
For checking a model before and after launch
Multilingual LLM data
Prompts and answers written by native speakers across the 50 most widely spoken languages. See translation and localization.
For models used outside English
Indian language LLM data
Hindi, Tamil and many more, including the mixed Hindi and English people actually type. See Indian language data.
For assistants for users in India
Where it's used
What teams build with it
A few of the jobs LLM data ends up doing. These are typical projects, not a client list. For raw text, the text data page covers collection and labeling.
Instruction-following assistants
Prompt and answer pairs that show a model what a helpful reply looks like.
Answers users prefer
Human preferences your team feeds into RLHF or a similar method.
Support and banking chat
Multi-turn conversations in the tone your customers expect.
Release checks
Rated answers across topics and languages, before and after launch.
Models for global users
Native-written data in the languages your users speak.
Assistants for India
Hindi, Tamil and mixed-script data from native speakers.
Before anyone writes a prompt
Four things to pin down first
Every LLM data project we've seen goes smoother when these are written down on day one, because they're what the writers and raters will argue about later.
What a good answer is
Tone, length, format and when the model should decline. Put examples of good and bad answers in the guideline.
How judgments are recorded
A simple pick, a full ranking or a score, and whether raters write a reason.
Topics and expertise
Which subjects the data has to cover, and how much expertise the writers need for each.
Languages and scripts
Which languages, whether answers are written natively or translated and checked, and how mixed-language messages are handled.
How it works
From your guideline to data your team can train on
One named person owns the project. Each step below links to the service behind it.
Brief
The kind of data, topics, languages and rough volume.
NDA first
Signed before any prompt or data moves.
Guideline
Written with you, with good and bad examples.
Writing
People from our network write prompts and answers.
Judging
Answers compared, put in order or rated, then reviewed.
Evaluation
Held-back prompts check your model after training.
Languages
LLM data beyond English
We cover the 50 most widely spoken languages in the world and go deep on Indian ones, well past Hindi and English. Answers are written by native speakers, not left as machine translation.
Glossary
LLM data terms, decoded
The words that come up on an LLM data project. The full AI training data glossary has the rest.
- Large language model (LLM)
- A model trained on large amounts of text to read and write language, like the ones behind chat assistants.
- Instruction tuning
- Training a model on prompts paired with good answers, so it learns to follow instructions. Also called supervised fine-tuning.
- RLHF
- Reinforcement learning from human feedback: tuning a model with human judgments of which answers are better.
- Preference data
- Records of which answer people preferred when shown two or more answers to the same prompt.
- Reward model
- A model trained on preference data to score answers the way people would, used during RLHF.
- Rubric
- The written criteria raters use to score an answer, like accuracy, tone and whether it followed the prompt.
- Multi-turn conversation
- A dialogue with several back-and-forth turns, used to teach a model to keep track of a thread.
- Red-teaming
- Writing difficult or adversarial prompts on purpose, to find where a model gives unsafe or wrong answers.
Related work
Services behind LLM data
The same team does all of these. You can also compare all eight services.
Data collection
People who write prompts and answers.
Data annotation and labeling
Answers compared, rated and reviewed.
Data and model evaluation
Your LLM's answers checked against targets.
Conversational AI data
Multi-turn dialogue for chat assistants.
Translation and localization
Native checks in every language.
FAQ
Questions about LLM training data
What people ask us about LLM data. The services FAQ covers how the services fit together.
What is LLM training data?
The text a large language model learns from and gets checked against: prompts paired with good answers, comparisons that say which answer is better, multi-turn conversations, and test sets. For a model that has to follow instructions, most of the useful data is written or judged by people.
What is RLHF data?
RLHF stands for reinforcement learning from human feedback. The data behind it is usually preference data: people read two or more answers to the same prompt and pick or rank the better one, often with a short reason. Your team then uses those judgments to train a reward model and tune the LLM.
What is the difference between instruction tuning data and preference data?
Instruction tuning data shows the model a prompt and one good answer to copy. Preference data shows it several answers and says which one people liked more. Teams often start with the first and add the second.
Do you train or fine-tune the LLM?
No. We provide the data, and your team trains or fine-tunes your own model with it. We can also check your model's answers afterwards, as part of our evaluation work.
Which languages can you cover?
The 50 most widely spoken languages in the world, with deep coverage of Indian languages, well past Hindi and English. Our Indian language data page has more on that side.
How do you keep human-written answers consistent?
Everyone works from the same written guideline, a second person reviews a share of each batch, and disagreements go back into the guideline as worked examples. That's the same loop we use for any labeling work.
Can we try a small sample first?
Yes. Send a short brief and we'll put together a small sample free before you commit.
Does our data stay private?
Yes. The NDA is signed before any data moves, we use your prompts and data only for the scope we agreed, and only the people on your project can see them. It's returned or deleted when you ask.
Start with a free sample
Pick the kinds of LLM data, add the languages and a rough volume, and send it. We'll put together a small sample free before you commit.
- Your named contact replies within one business day.
- We sign an NDA before any data changes hands.
- We prepare your free sample, then agree the full scope.
This opens your email app with the brief filled in. Nothing is sent until you press send. You can also use the contact page.