Practical Guidance
What to curate first if you are new to training datasets.
What stage are you improving?
Each stage targets a different capability. Pre-training builds knowledge. Fine-tuning teaches assistant behavior. Preference training improves answer quality. RL trains multi-step actions. Pick the stage where your model is weakest.
What unit of training example fits that goal?
The shape of your training examples follows from the stage. Pre-training uses raw document chunks. Fine-tuning uses prompt-answer pairs or multi-turn chats. Preference learning uses ranked or scored answers. RL uses action-observation trajectories.
What behavior do you want to teach?
Be specific. 'Make the model better' is too vague. Target a concrete behavior: follow formatting instructions, cite retrieved sources, call the right API, refuse unsafe requests, or answer domain questions accurately.
What label do you actually have?
Your labels constrain which training methods you can use. If you only have raw text, you can pre-train. If you have ideal answers, you can fine-tune. If you have rankings or scores, you can do preference training. If you have success/failure signals, you can do RL.
How will you measure success?
Build a held-out evaluation set before you start training. Without one, you cannot tell whether your data improved the model, hurt it, or made no difference.
Bottom Line
All training datasets are not the same. They differ along three axes:
- Stage — pre-training, fine-tuning, preference, RL, safety
- Example shape — document, prompt-response, chat, ranking, trajectory
- Behavior target — knowledge, helpfulness, style, safety, grounding, tool use, planning