Choose a post-training path

Every path starts from a released Cosmos3 checkpoint; the sections below differ by the training signal you bring.

You have	Method	Where
Paired examples: prompt → video, video → text answer, observation → action	SFT	SFT
A teacher model you want to make faster (fewer steps) or causal	Distillation	Distillation
A reward function, verifier, simulator or preference pairs	RL	RL

SFT

Supervised fine-tuning runs on the native trainer. Each recipe is a TOML file under examples/toml/sft_config/; most have a paired examples/launch_sft_*.sh that runs the trainer under torchrun.

python -m cosmos_framework.scripts.train --sft-toml=examples/toml/sft_config/<recipe>.toml

• SFT recipes: Generator SFT for Cosmos3-Nano, Cosmos3-Super (LoRA) and Cosmos3-Edge, Reasoner SFT for all three, plus checkpoint conversion and export.
• Your own data: JSONL datasets and custom datasets.
• Robot actions from Cosmos3-Nano: DROID policy (reproduces Cosmos3-Nano-Policy-DROID), DROID forward dynamics and LIBERO-10 policy.
• Transfer-Auto: a four-phase RGB/LiDAR curriculum from Cosmos3-Nano, run with cosmos_framework.scripts.train_experiment instead.

Distillation

• Sim-Auto: three stages from Transfer-Auto, TF → TF-dCM → SF-DMD. To adapt Sim-Auto to your domain, adapt Transfer-Auto first and then distill, rather than running SFT on the released Sim-Auto checkpoint; this preserves its few-step generation.
• Sim-Transfer: a bidirectional Stage 0 teacher, then Stages 1–3: teacher forcing, four-step distillation and self-forcing.
• Sim-Bimanual: Stages 0–3 from Cosmos3-Nano. Check the training data before loading weights.
• Super four-step (DMD2): Cosmos3-Super-Text2Image or Cosmos3-Super-Image2Video as the teacher, the matching -4Step checkpoint as the initial student. A six-iteration functional recipe, not a production reproduction.

RL

RL runs on the cosmos_framework.posttrain runtime in its own environment: the Framework base plus a locked dependency set. Start with the overview and the run guide.

• Generator: DPO learns from preference pairs collected in advance; DiffusionNFT generates and scores its own samples with a reward. The run-guide recipes start from Cosmos3-Nano.
• Reasoner: GRPO, PPO and OPD (on-policy distillation from a frozen teacher). The run-guide recipes use public Qwen models.

To start the Reasoner side from Cosmos3-Nano or Cosmos3-Super, export the checkpoint as a Qwen3-VL Hugging Face directory. For Cosmos3-Super, add --vlm-model-name Qwen/Qwen3-VL-32B-Instruct.

python -m cosmos_framework.scripts.convert_model_to_vlm_safetensors \
    --checkpoint-path Cosmos3-Nano \
    -o examples/checkpoints/Cosmos3-Nano-VLM

Joint PPO also trains GR00T-N1.7 and π0.5 robot policies on LIBERO-Spatial; see the GR00T-N1.7 and π0.5 guides. The runtime also has an SFT objective for Qwen text and vision-language models.

AutoML (preview)

AutoML searches Reasoner SFT hyperparameters against held-out evaluation; its GEPA Auto-Prompter optimizes evaluation prompts without training model weights. It requires the nvidia-automl-core wheel, which is not yet published to a package index. See AutoML.