Source-backed overview
Mercor
An expert-data platform supplying frontier training data, evaluation work, and RL environments.
- Subsector
- Expert data platforms
- Primary capability
- Expert-generated data and evaluation
Independent industry field guide · v0.1
How AI systems learn after the internet.
AI Training Atlas maps the data, expert work, evaluations, and interactive environments that turn general-purpose models into useful systems.
3 / 10 planned
13 core terms
1 foundation
Claim-labeled
01 · Industry foundations
Training is a lifecycle, not a single event.
A model moves through several distinct systems. Each stage creates different needs for data, judgment, infrastructure, and evaluation.
- 01
Pretraining
Build a base model from broad data and predictive objectives.
- 02
Post-training
Shape behavior with demonstrations, preferences, outcomes, and practice.
- 03
Evaluation
Measure capabilities, trade-offs, and failures under defined conditions.
- 04
Inference & operation
Run the model in products or agent systems with tools and environments.
02 · Company landscape
A market forming around model improvement.
The initial landscape focuses on expert-generated data and agent environments. Profiles distinguish sourced facts from unavailable information.
Partial public data
AfterQuery
An applied research lab turning expert reasoning and real workflows into model training data.
- Subsector
- Expert data platforms
- Primary capability
- Expert reasoning data and agent training
Early-stage public data
Fleet
An applied research lab building high-fidelity training environments for AI agents.
- Subsector
- RL environment providers
- Primary capability
- High-fidelity agent environments
03 · Working vocabulary
Learn the language of the stack.
13 concise definitions connect training methods to the companies and lifecycle stages where they matter.
PretrainingThe broad first phase in which a model learns patterns from a very large dataset.Post-trainingTraining after pretraining that shapes a model into a useful, reliable assistant or agent.Supervised fine-tuningTraining on curated examples that demonstrate desired inputs and outputs.Data annotationThe process of adding labels, judgments, corrections, or other structure to data used for training or evaluation.Preference dataHuman or model judgments that compare outputs or score their quality.RLHFReinforcement learning from human feedback used to align model behavior with human preferences.
04 · Ecosystem map
Clusters, not false precision.
The map groups companies by primary emphasis without implying market share or exact competitive distance.
Active subsector
Expert data platforms
Active subsector
RL environment providers
05 · Research watch
Signals worth following.
A small curated layer separates consequential developments from the future external feed.