Home / Label types / LLM data
RLHF and preference data
Human raters compare, rank and score model responses, so your model learns what a good answer looks like to real people.
PromptExplain UPI to my grandmother in two sentences.
Response AUPI is a real-time payment system developed by NPCI that enables inter-bank transactions via mobile devices.
Response BUPI lets you send money from your phone to anyone, instantly. It is like handing over cash, but through your bank.
Helpful
Accurate
Safe
Preferred: BRationale: plain language, fits the audience, still correct.
RaterCompared
ReviewerChecked
QAApproved
What we label
- Pairwise preference: which answer is better
- Ranking several responses to one prompt
- Rubric scores for helpfulness, accuracy and safety
- Written rationales for every choice
- Rewritten ideal answers for fine-tuning
Where it is used
- Chatbots and AI assistants
- English, Hindi and Hinglish models
- Safety and refusal tuning
- Customer support bots
How we check quality
- Calibration rounds on gold examples before live work
- Agreement between raters measured on shared items
- Rationales checked for real reasons, not filler
- Reviewer audits on a sample of every batch
Delivery formats
We deliver in the format your training pipeline already uses. Need something else? Send us a sample file and we will match it.
JSONLCSVHugging Face datasetYour own schema
See rlhf and preferences on your own data
Send a small sample and your guidelines. The pilot is free.