Skip to content
Karmyug

Preference ranking pairs

Pairwise preferences with a written reason on every comparison.

8 Indian languagesLLM alignment

Overview

Preference data for RLHF and DPO where the rater must justify the choice in writing. Disagreement rate is published per language, which matters more than the headline agreement number when you are weighting a training mix.

Specification

Format
JSONL
Rationale
Required

Sample

Illustrative sample. It shows the structure, fields and metadata of this dataset with invented values, not real recordings or records. A real sample pack is sent on request.

Sample records for Preference ranking pairs
idlanguageturnsdomainspeaker_rolesexcerptlabelreviewer_agreementqa_status
tx_0001English6LLM alignmentwriter, reviewerCustomer reports a delayed refund and asks for a status update.resolvedagreeaccepted
tx_0002English12LLM alignmentwriter, reviewerAgent confirms identity and explains what happens next.escalatedagreeaccepted
tx_0003English4LLM alignmentwriter, reviewerUser asks a follow-up that depends on the previous turn.resolvedsplit, adjudicatedaccepted
tx_0004English9LLM alignmentwriter, reviewerCustomer switches topic and the agent re-confirms the goal.clarifiedagreeaccepted
tx_0005English7LLM alignmentwriter, reviewerConversation closes with a summary and a confirmation.resolvedagreeaccepted

Quality assurance

  • Rubric qualification
  • Hidden gold items
  • Speed anomaly detection

Want a real sample before you commit?

The sample above shows the structure. We also send a representative slice that includes the hard cases, plus the QA report and licence terms. No call required.