- Home
- Training data
- Human Feedback & Safety
Human feedback and safety data.
Preference, demonstration, and safety data from generalist and expert raters, produced to your guidelines for fine-tuning, reward modeling, and safety work.
What we can supply
- Preference pairs for RLHF and DPO
- Rubric scoring for reward and judge models
- Demonstrations and instruction-following responses
- Multi-turn dialogue
- Expert Q&A in regulated domains
- Red-team and adversarial prompts
- Content-moderation labels
- Factuality and hallucination labels
Labels and enrichment
- Preferences with rationale
- Per-axis scores
- Gold responses
- Severity
- Policy mapping
- Claim and evidence spans
- Rater metadata
Formats and delivery
JSONL with prompts, responses, labels, and rater metadata. Delivered to AWS S3, Google Cloud Storage, Azure, or your own storage, with a datasheet and provenance records.
Common uses
- Reward modeling
- RLHF and DPO
- Supervised fine-tuning
- Safety training
- Guardrails
- Hallucination detection
Questions
Can you work to our annotation guidelines?
Yes. We run calibration rounds against your guidelines and rubric before production starts.
Who are the raters?
Generalist and credentialed expert raters, matched to the task, with agreement measured and reported.
How is harmful content handled?
Red-team and moderation work runs with wellbeing safeguards for raters and strict access controls on outputs.
Own data like this? License it through ModelCore