Conversations worth learning from.
Multi-turn dialogue, speech and human feedback, ranked and reviewed by people who know the domain, for assistant training and evaluation.
Available as- 01 · Dataset
- Multi-turn dialogue, intent-labelled
- 02 · Environment
- Multi-turn evals, rubric-graded
Licensed, stocked, ready to ship.
Plus custom collection for what doesn't exist yet.
Contact centre, media and voice workflows
Examples prepared for advanced models
Support, assistants, enterprise and domain workflows
Independent passes on preference data
SFT, RLHF and DPO
Continuous annotation and review
Dialogue, ranked by experts.
Every conversation carries intent, context and the human judgment that tells a model which answer is better.
- Multi-turn dialogue
- Support, assistant and enterprise conversations with full context.
- Instruction data
- Prompts and expert-written responses for supervised fine-tuning.
- Preference pairs
- Chosen and rejected responses with rubric scores for RLHF and DPO.
- Intent & sentiment
- Labels for routing, tone and outcome.
- Safety review
- Red-team transcripts and adversarial prompts with graded responses.
- Speech dialogue
- Spoken conversations with transcripts for voice assistants.
Preference pair. Values are illustrative.
From live chat to ranked pairs.
One support conversation, through the five stages every dialogue goes through before it ships.
Source
Conversations come from licensed support logs and from sessions written by people who know the domain, never scraped from forums.
- Who
- Licensed logs + experts
- Gate
- Known source for every conversation
De-identify
Names, addresses and order numbers are replaced, so the conversation keeps its meaning without the customer.
- Who
- Model + reviewer
- Gate
- Every replacement human-checked
Candidates
Where a turn matters, more than one assistant answer is written or collected, so reviewers have real alternatives to compare.
- Who
- Writers
- Gate
- At least two answers to rank
Expert ranking
Experts rank the answers, score them against a rubric and write why the rejected one fails. Disagreements go to a third reviewer.
- Who
- Domain experts
- Gate
- Low agreement is redone
Package
Ships as message threads, preference pairs and rubric scores, in the schema your trainer reads.
- Who
- Delivery
- Gate
- Rubric shipped with the data
Built for frontier model tasks.
Scope, review criteria and delivery format are agreed per engagement, from the first sample to the final eval.
- Assistant SFT
- Expert-written responses across domains.
- RLHF & DPO
- Ranked pairs with reviewer agreement metrics.
- Customer support agents
- Real support dialogue with intents and outcomes.
- Safety & red teaming
- Adversarial prompts with graded responses.
- Dialogue evaluation
- Rubric-based eval sets for assistants.
What buyers ask us first.
Something else? Ask a specialist →
01What is conversational AI data?
Text dialogues, speech recordings, transcripts, intent and sentiment labels, and human preference feedback used to train and evaluate chatbots, voice assistants and other dialogue systems.
02What is RLHF data?
Human comparisons and ratings of model responses used as preference data. Training with it helps assistants give more useful, relevant and safer answers.
03How do you keep rankings consistent?
Reviewers are screened for the domain, trained on your rubric and calibrated in rounds. Every batch ships with agreement metrics.
04What does a preference record include?
The prompt, the chosen and rejected responses, the reason the rejected one lost, rubric scores, who wrote and reviewed it, and the agreement between two independent reviewers. The rubric version each pair was made under ships with the data.
Same standard in every domain.
Private code
Repositories, commits, pull requests and review history.
- 01 Dataset
- 3B+ lines with commit, PR and review history
- 02 Environment
- Agentic engineering tasks, engineer-graded
Healthcare
Clinical notes, records and imaging, de-identified and expert-reviewed.
- 01 Dataset
- 10M+ de-identified, patient-linked records
- 02 Environment
- Clinical case tasks, clinician-graded
CAD & 3D
Drawings, assemblies and meshes with validated geometry.
- 01 Dataset
- 5M+ typed, native-format CAD files
- 02 Environment
- Part classification and rebuild checks
Audio & speech
Synchronized audio with transcripts and event labels.
- 01 Dataset
- 5.7M+ transcribed, diarized assets
- 02 Environment
- Transcription evals per language and accent
Egocentric video
First-person footage with action and object labels for physical AI.
- 01 Dataset
- 200K+ hrs synced video, IMU, action labels
- 02 Environment
- Action and manipulation benchmarks
Company operations
Workflows, approval trails and legal documents for enterprise agents.
- 01 Dataset
- 52+ companies' ordered approval trails
- 02 Environment
- Agent tasks over approvals, operator-graded
See real conversational data first.
A representative sample under NDA, in the format your pipeline reads. Run it through your own evals first.
- Sample records, as delivered
- Datasheet and schema
- Review criteria
- JSONL (messages) · Preference pairs · Rubric scores
{"prompt":"Refactor this handler to avoid
the N+1 query.",
"chosen":"Use select_related('org') and batch…",
"rejected":"Wrap the loop in a try/except…",
"rubric":{"correct":5,"safe":5,"clear":4},
"reviewers":2,"agreement":0.94}