Cosmos-Reason1 Post-Training Hugging Face Example

This package provides a minimal Cosmos-Reason1 post-training example using the Hugging Face datasets format. You should first read the full post-training example, see Cosmos-Reason1 Post-Training Full.

Setup

Install

Prerequisites:

• Setup

Install the package:

cd examples/post_training_hf
just install
source .venv/bin/activate

Example

Download the Nexar collision prediction dataset:

./scripts/download_nexar_collision_prediction.py data/sft --split "train[:10]"

Run SFT:

cosmos-rl --config configs/sft.toml scripts/custom_sft.py

The full config is saved to outputs/sft/config.toml.