Training and evaluation data for language and multimodal models.
From domain corpora to expert-written examples, preference judgments, and agent trajectories. We scope each dataset around the model behavior you want to change and the way you will measure it.
Who it is for
Teams we work with
- Foundation-model and research labs
- Applied AI teams fine-tuning for a domain
- Enterprise AI teams preparing internal knowledge
- Evaluation and safety teams
Objective
Improve or measure specific model behavior: domain knowledge, instruction following, judgment quality, tool use, or robustness to difficult inputs.
Data categories
What AI Model Data covers
Pretraining & domain adaptation
Licensed and cleaned corpora in the domains and languages your model needs to know.
Instruction & supervised fine-tuning
Instruction–response pairs written to your style guide and task mix.
Expert-authored examples
Worked examples from reviewers with relevant professional background, scoped per project.
Preference & rubric ratings
Pairwise comparisons and rubric scores with written rationales for alignment work.
Evaluation & adversarial sets
Held-out test cases with expected outputs, grading notes, and difficult edge cases.
Enterprise knowledge preparation
Your documents chunked, cleaned, de-identified, and structured for retrieval or tuning.
Agent tool-use trajectories
Multi-step traces of plans, tool calls, observations, and outcomes.
Synthetic augmentation
Model-assisted generation where appropriate, always labeled and human-reviewed.
Know what you are getting
Written, judged, and recorded data serve different purposes.
We label every record by how it was produced so your team knows what it can be trusted for.
Authored
Examples written by people to a specification: instructions, responses, test cases.
Best for teaching a target behavior.
Judged
Human ratings of existing outputs: comparisons, rubric scores, rationales.
Best for alignment and reward modeling.
Recorded
Traces of real or simulated workflows: tool calls, observations, final results.
Best for agent training and replay evaluation.
Illustrative samples
Example records
HX-SFT-0318 Illustrative
- Instruction
- Summarize the result below for a patient newsletter in three sentences. Keep the numbers, avoid jargon, and do not imply the treatment is approved.
- Input
- Phase II, n=120. Primary endpoint: 12-week reduction in HbA1c. Treatment arm −0.9 percentage points vs −0.3 placebo (p<0.01). Mild GI events in 14% of treated participants.
- Response
- In a mid-stage study of 120 people, those who took the study drug saw their long-term blood sugar measure fall by 0.9 points over 12 weeks, compared with 0.3 points for those on a placebo. About 1 in 7 people taking the drug had mild stomach side effects. The drug is still being studied and is not yet approved.
- Style tags
- plain-language · 3 sentences · no approval claims
- Review
- Numbers verified against input · approved
Clinical-trial summary instruction
Instruction–response pair written to a style guide, with a reviewer check on factual fidelity.
- Pillar
- Model Data
- Modality
- Text & Documents
- Use case
- Supervised fine-tuning & instructions
- Method
- Expert-authored
- Instruction
- Reference response
- Style tags
- Fidelity review
Specification
What we agree before production
- Task & domain
- The capability being trained or tested, and the subject area it applies to.
- Language
- Languages and registers included, stated per dataset.
- Record structure
- Field schema for each record, with an example in JSON.
- Annotation or review method
- Who wrote or judged each record, against which guideline, and how disagreements were resolved.
- Dataset split
- Train, validation, and held-out evaluation splits, kept separate from the start.
- Delivery format
- Usually JSONL or Parquet, with a schema file and data card.
- Usage scope
- Permitted uses and restrictions, agreed in writing for each project.
Quality & governance
- Written guidelines with worked examples before production begins
- Calibration rounds and inter-annotator agreement checks
- Contamination checks between training and evaluation splits
- De-identification of personal data where required
Example deliverable
- model-data-delivery/
- data/
- train.jsonl
- validation.jsonl
- eval_heldout.jsonl// kept separate
- schema.json// field definitions
- guidelines.pdf// authoring + review rules
- review/
- agreement_report.json
- qa_sample.csv
- DATA_CARD.md// provenance, intended use, limits
Have a specific dataset in mind?
Share the task, scale, and timeline. We will reply with questions and a proposed scope.