trl-training

Trainieren und Feinabstimmen von Transformer-Sprachmodellen mit TRL (Transformers Reinforcement Learning). Unterstützt SFT, DPO, GRPO, KTO, RLOO und Reward-Modell-Training…

npx skills add https://github.com/huggingface/trl --skill trl-training

TRL

Each method pairs a *Trainer class with a *Config dataclass. Configs extend transformers.TrainingArguments, so all of its arguments work in any trainer config.

TrainerDataset type
SFTTrainerlanguage modeling or prompt-completion
DPOTrainerpreference (chosen/rejected pairs)
GRPOTrainerprompt-only + reward function(s)
DistillationTrainerprompt-only + a teacher model (on-policy distillation)
KTOTrainerunpaired preference (per-sample bool label)
RewardTrainerpreference (chosen/rejected pairs); trains a scalar reward model, not a policy

Many more trainers (OnlineDPO, ORPO, CPO, GKD, …) live in trl.experimental with unstable APIs: https://huggingface.co/docs/trl/experimental_overview

from datasets import load_dataset
from trl import SFTConfig, SFTTrainer

trainer = SFTTrainer(
    model="Qwen/Qwen2.5-0.5B",  # model ID or a PreTrainedModel instance
    args=SFTConfig(output_dir="Qwen2.5-0.5B-SFT"),
    train_dataset=load_dataset("trl-lib/Capybara", split="train"),
)
trainer.train()

Pass model as a string and route loading kwargs through model_init_kwargs (e.g. {"dtype": "bfloat16", "attn_implementation": "kernels-community/flash-attn2"}) instead of calling from_pretrained yourself. The tokenizer/processor is inferred from the model; pass processing_class only when it differs. For LoRA, pass peft_config=LoraConfig(...).

Dataset formats

Conversational: {"messages": [{"role": ..., "content": ...}]} (language modeling) or {"prompt": [...], "completion": [...]}. The chat template is applied automatically — never apply it yourself. Extra columns are allowed; GRPO forwards them to reward functions. Reference: https://huggingface.co/docs/trl/dataset_formats

SFT: the fields that matter

SFTConfig(
    max_length=1024,        # truncation length; None disables truncation
    packing=True,           # pack sequences into max_length blocks: fewer pad tokens, higher throughput
    padding_free=True,      # flatten batch, no padding; requires FlashAttention; implied by packing
    use_liger_kernel=True,  # fused Liger kernels, reduces peak memory
    assistant_only_loss=True,  # loss only on assistant turns (conversational datasets)
)

GRPO: online RL

def reward_len(completions, **kwargs):
    return [-abs(20 - len(c[0]["content"])) for c in completions]

trainer = GRPOTrainer(
    model="Qwen/Qwen2.5-0.5B-Instruct",
    reward_funcs=reward_len,  # or a list; rewards are summed
    args=GRPOConfig(output_dir="Qwen2.5-0.5B-GRPO", max_completion_length=512),
    train_dataset=load_dataset("trl-lib/DeepMath-103K", split="train"),
)

Reward functions are called with keyword arguments prompts, completions, completion_ids, trainer_state, plus every extra dataset column — accept **kwargs for the ones you ignore. Return list[float], one reward per completion. With conversational data, completions is a list of message lists, not strings.

The generation batch is per_device_train_batch_size × num_processes × steps_per_generation (or set generation_batch_size directly) and must be divisible by num_generations (default 8). Generation is the usual bottleneck — enable vLLM with use_vllm=True: vllm_mode="colocate" shares the training GPUs (size with vllm_gpu_memory_utilization); vllm_mode="server" uses a separate trl vllm-serve --model <model_id>.

AsyncGRPOTrainer (trl.experimental.async_grpo) implements the same algorithm with generation decoupled from training: a background worker streams completions from a vLLM server while the training loop consumes them, so the two overlap instead of alternating.

CLI

Flags mirror the config fields: trl sft --model_name_or_path Qwen/Qwen2.5-0.5B --dataset_name trl-lib/Capybara. YAML via --config; distributed presets via --accelerate_config zero3 (Python scripts: accelerate launch train.py).

Mehr Skills von huggingface

sync-models
huggingface
Synchronisiere die Modellkonfiguration von chat-ui mit dem HuggingFace-Router — füge Beschreibungen für neue Modelle hinzu, kennzeichne reasoning-fähige Modelle, aktiviere Artefakte für Modelle mit 32B+…
custom-blocks
huggingface
Use when the user has written (or wants to write) a `ModularPipelineBlocks` subclass in a local Python file and needs to package it into a Hub-uploadable…
self-review
huggingface
Use before opening a PR, or whenever asked to self-review a diffusers contribution. Applies the same rubric as the `@claude` CI (checks the diff against…
hf-cloud-sagemaker-production-defaults
huggingface
Erstellen Sie einen SageMaker-Endpunkt (Echtzeit oder asynchron) mit automatischer Skalierung, CloudWatch-Alarmen und standardmäßig aktiviertem Tagging. Verwenden Sie diese Fähigkeit, wenn Sie kurz davor sind, einen … zu erstellen.
hf-cloud-serving-image-selection
huggingface
Wählen Sie den richtigen Serving-Container für eine SageMaker-Modellbereitstellung aus und finden Sie dessen aktuelles Image-URI. Verwenden Sie diese Fähigkeit, wenn Sie ein Modell auf einer… bereitstellen möchten.
Hugging Face Cli
huggingface
Execute Hugging Face Hub operations using the `hf` CLI. Use when the user needs to download models/datasets/spaces, upload files to Hub repositories, create repos, manage local cache, or run compute jobs on HF infrastructure. Covers authentication, file transfers, repository creation, cache operations, and cloud compute.
Hugging Face Datasets
huggingface
Erstellen und verwalten Sie Datensätze auf dem Hugging Face Hub. Unterstützt das Initialisieren von Repos, das Definieren von Konfigurationen/System-Prompts, das Streamen von Zeilenaktualisierungen und SQL-basierte Abfragen/Transformationen von Datensätzen. Entwickelt, um zusammen mit dem HF MCP-Server für umfassende Dataset-Workflows zu arbeiten.
Hugging Face Evaluation
huggingface
Evaluierungsergebnisse in Hugging Face-Modellkarten hinzufügen und verwalten. Unterstützt das Extrahieren von Evaluierungstabellen aus README-Inhalten, das Importieren von Bewertungen aus der Artificial Analysis API und das Durchführen benutzerdefinierter Modellevaluierungen mit vLLM/lighteval. Funktioniert mit dem model-index-Metadatenformat.