Two things we ship.
Datasets your model learns from. Environments where it proves it. Seven verticals, two outputs.
Pick the domain you need.
Data licensing, custom data collection and expert data annotation, in every vertical. Each has its own sources, risks and reviewers.
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
Conversational
Multi-turn conversations, ranked and reviewed for assistant training.
- 01 Dataset
- Multi-turn dialogue, intent-labelled
- 02 Environment
- Multi-turn evals, rubric-graded
Shipped in your pipeline's format.
Pick a deliverable. See the folder you receive, one example record from it, and the checks every record passed.
- corpus_healthcare_v2/
- └ shards/
- └ shard_0000.jsonl
- └ shard_0193.jsonl
- └ …
- └ datasheet.md
- └ schema.json
- └ provenance.csvper record
- └ dedup_report.json
- Format
- JSONL · Parquet · your schema
- Unit
- Document, with source and licence
- Docs
- Datasheet and schema
- Rights on every sourceLicence and consent documented before collection.
- Identifiers removedAutomatic pass, then human spot checks.
- Duplicates removedExact and near-duplicate filtering across shards.
- Quality filteredBoilerplate, broken encodings and junk dropped.
- post_training_code_v1/
- └ sft/
- └ train.jsonl
- └ preference/
- └ pairs.jsonl
- └ rubric_v3.md
- └ agreement.csvper batch
- └ datasheet.md
- Format
- JSONL · chat or pair schema
- Unit
- Example or ranked pair
- Reviewers
- Domain experts
- Written by practitionersEngineers, clinicians, linguists: not general crowds.
- Expert reviewEvery batch is reviewed by a domain specialist.
- Agreement trackedBatches below the agreed threshold are redone.
- Rubric includedYou get the rubric each label was made under.
- env_agentic_eng_v1/
- └ tasks/
- └ task_0412/
- └ repo.tar
- └ task.yaml
- └ tests/
- └ grader/
- └ grade.py
- └ env_card.md
- └ baselines.json
- Grader
- Programmatic + expert rubric
- Unit
- Task, with reference solution
- Domains
- Code, clinical, ops, CAD
- Realistic workTasks come from real repositories, records and workflows.
- Grader verifiedGraders pass the reference solution and fail broken ones.
- No leakageChecked against public repositories and benchmarks.
- Difficulty calibratedTasks spread across easy, medium and hard.
- eval_clinical_reasoning_v1/
- └ items.jsonl
- └ answer_key.jsonlheld out
- └ rubric.md
- └ scoring.py
- └ baselines.json
- Format
- JSONL + scoring script
- Unit
- Item, with rubric
- Graders
- Domain experts
- Written, not scrapedEvery item is new, so models haven't seen it.
- Answer key held outKept separate from anything you train on.
- Expert gradersScored by domain specialists against a rubric.
- Contamination checkSearched against public data before release.
Environments and evals, explained.
01What is an RL environment, and what do you deliver?
An RL environment is a set of tasks an agent can attempt, with a grader that scores each attempt automatically. Ours ship as task suites built from real repositories, records and workflows, each with a programmatic grader, an expert rubric and a human reference solution the grader was checked against.
02What's the difference between a dataset and an environment?
A dataset is what a model learns from; an environment is where it proves it. Both start from the same licensed, cleaned records. For an environment, domain experts add tasks and graders in a second pass, and evaluation items are held out from anything you train on.
03How do you keep eval sets out of training data?
Eval items are written new rather than scraped, the answer key is kept separate from training deliveries, and every item is searched against public data and benchmarks for contamination before release.
04Do you build red-team sets?
Yes. Adversarial prompts with graded responses, written and scored by domain experts, show where a model is unsafe or easily misled. They ship with the rubric they were graded under.
Pick a domain, see samples.
Three choices. We write the email; a specialist for that domain replies with availability, rights and a sample plan.
- 01A reply from a specialist in that domain
- 02A representative sample under NDA
- 03Its datasheet, schema and review criteria