SFT data and prompt writing
Prompts and ideal responses written by native speakers and subject experts for supervised fine-tuning: realistic questions, model answers in your style, multi-turn conversations and system prompts. Every example is reviewed before it reaches your training set.
Primeiro, estacione em um lugar plano e seguro, ligue o pisca-alerta e puxe o freio de mão.
Depois, siga estes passos:
- Afrouxe os parafusos da roda antes de levantar o carro.
- Posicione o macaco no ponto indicado no manual.
- Troque o pneu e aperte os parafusos em cruz.
- We write
- Prompts, ideal responses, multi-turn conversations and system prompts
- Writers
- Native speakers, with subject experts for specialist topics
- Style
- Your style guide, tone, length and formatting rules
- Output
- JSONL in your chat format, with metadata for each example
What SFT data is
Supervised fine-tuning (SFT) data is a set of example prompts paired with the responses you want a model to give. Fine-tuning on good examples teaches a model your tone, format and level of detail, and how to handle tasks it gets wrong today. The quality of each example matters more than the number of them.
Translated SFT data teaches a model to sound translated. We write examples natively in each language, so the model learns how people there ask questions and expect to be answered. For preference rankings, see RLHF data.
What we write
To your task list, style guide and language mix.
Prompt and response pairs
Realistic user requests and the ideal answer, across the tasks your model must handle.
Multi-turn conversations
Conversations with follow-ups, corrections and changes of topic, written as one coherent dialog.
Expert responses
Answers in medicine, law, finance, science and coding from writers with that background.
Reasoning examples
Step-by-step worked solutions for math and logic tasks, checked for correct final answers.
Refusals and safe answers
Examples of declining harmful requests politely and helping with the safe part.
Prompt writing
System prompts and test prompts written and tested for your product.
How an SFT data project runs
You approve sample examples before volume writing starts.
Task list and style
The tasks, languages, tone and formatting rules for responses.
Samples
A small set of examples per language for your team to approve.
Writing
Writers produce examples, balanced across tasks and difficulty.
Review
A second reviewer checks accuracy, style and language on every example.
Second review QADelivery
JSONL in your chat template, with task, language and writer metadata.
Related AI data services
Questions about SFT data
What is SFT data?
SFT data is the set of prompts and ideal responses used for supervised fine-tuning, the step where a pretrained language model learns to follow instructions and answer in a particular style.
Is SFT data written or translated?
Written. We write prompts and responses natively in each language, because translated examples teach a model to sound translated. Translation is available if you specifically need parallel examples.
Who writes the responses?
Native-speaker writers trained on your style guide, with subject experts for specialist topics such as medicine, law or coding. A second reviewer checks every example.
What format do you deliver?
JSONL in the chat template your training pipeline uses, with metadata such as task type, language and difficulty for each example.
Request sample examples
Send your task list and style guide, or book a call. We write sample examples per language for you to approve.
- Phone
- +44 20 3290 0116
- Phone hours
- Weekdays 9:30 to 16:00, weekends 9:30 to 12:00, UK time