trl-jobs 0.1.0

TRL-specific job runner built on hfjobs

TRL Jobs

A convenient wrapper around hfjobs for running TRL (Transformer Reinforcement Learning) specific workflows on Hugging Face infrastructure.

Installation

pip install trl-jobs

Available Commands

SFT (Supervised Fine-Tuning)

Run SFT job with ease:

trl-jobs sft --model MODEL_NAME --dataset DATASET_NAME [OPTIONS]

Required Arguments

• --model: Model name or path (e.g., Qwen/Qwen3-4B-Base)
• --dataset: Dataset name or path (e.g., trl-lib/tldr)

Optional Arguments

• --flavor: Hardware flavor (default: t4-small)
• -d, --detach: Run job in background and print job ID
• --token: Hugging Face access token

Examples

trl-jobs sft \
    --model Qwen/Qwen3-4B-Base \
    --dataset trl-lib/tldr

Hardware Flavors

Common hardware flavors you can use:

• t4-small: NVIDIA T4 GPU (default)
• t4-medium: NVIDIA T4 GPU with more resources
• a10g-small: NVIDIA A10G GPU
• a10g-large: NVIDIA A10G GPU with more resources
• a100-large: NVIDIA A100 GPU

Authentication

You can provide your Hugging Face token in several ways:

1. Using huggingface-hub login: huggingface-cli login
2. Setting the HF_TOKEN environment variable
3. Using the --token argument

License

MIT License - see LICENSE file for details.