Feature request

TRL already documents the expected dataset type for each trainer in docs/source/dataset_formats.md, but that information is not exposed programmatically.

Today, downstream wrappers and validation layers have to hardcode trainer-name checks like if trainer_class is GRPOTrainer or maintain their own registry to figure out whether a trainer expects prompt-only, preference, prompt-completion, etc.

A small metadata API on trainer classes would solve this cleanly:

from trl import GRPOTrainer

GRPOTrainer.dataset_types == ("prompt-only",)

For trainers that support more than one type:

from trl import SFTTrainer

SFTTrainer.dataset_types == ("language-modeling", "prompt-completion")

Motivation

This would let wrappers, CLIs, and downstream libraries:

• validate datasets before trainer construction
• choose preprocessing paths without hardcoded class checks
• stay aligned with the dataset-type table already maintained in the docs

Why now

There seems to be recurring user and maintainer effort around trainer-specific dataset semantics and chat-template behavior, for example:

• [SFT] Can't apply_chat_template prompt-completion dataset. so, training result always bad. #3468
• Inconsistent apply_chat_template behaviour for multimodal dataset between SFT and GRPO #3915
• ♻️ Reuse multimodal message preparation from SFTTrainer in GRPOTrainer #3919
• Prompt-completion does not use correct labels when passing conversational format to SFT #4147
• Fix prompt-completion labeling with add_generation_prompt and warning #4201
• GRPOConfig.max_prompt_length does not truncate prompts for a conversational dataset with vllm (since v0.25.0) #4358

Those issues are not asking for this exact API, but they do show that trainer dataset semantics matter in practice and are currently discovered indirectly through docs or implementation details.

Proposed shape

Add a class attribute on trainers:

class _BaseTrainer(Trainer):
    dataset_types: tuple[str, ...] = ()

Then set it on the trainers already covered by the dataset-format table.

This would be metadata only, with no behavior change.