← All services
RLHF & Preference Data
Pairwise preference from people whose language you're aligning to.
A/B preference collection with structured reasons, built for reward modelling and preference optimisation in Darija, Arabic, French and code-switched contexts.
What we cover
Coverage and categories
Scope any combination below, or bring your own taxonomy — the platform is configured per project.
Pairwise comparison
Two candidate responses, one judgement, one reason set.
Reason taxonomy
More natural · more accurate · understands dialect · more relevant · safer · better French/Arabic usage.
Tie handling
Explicit Equal and Both bad options so weak pairs are not forced.
Free-text rationale
Optional written justification for every judgement.
Multi-rater redundancy
Configurable overlap to measure agreement and filter noise.
Deliverables
What you receive
- Pairwise judgements: A better · B better · Equal · Both bad
- Structured reason codes plus free-text rationale
- Annotator calibration and agreement metrics
- JSONL export ready for reward-model training
Outcomes
Reward models that prefer what Algerians actually preferReduced reward hacking on dialect promptsAuditable preference data with reasons attached
Typical specification
- Export
- JSONL / CSV / Parquet-ready
- Overlap
- Configurable per batch
- Throughput
- Thousands of pairs per week
- Quality gate
- Gold-standard items seeded per batch