Modelpublic

cy0307/ropedia-cosmos3-super-forward-dynamics-lora-128ep

sourceHugging Faceotherupdated 3mo agoView on Hugging Face

1likes

Model Card

Ropedia Xperience-10M Cosmos3-Super Forward-Dynamics LoRA

This model repo contains the weight-bearing LoRA adapter tensor file for the verified Cosmos3-Super forward-dynamics branch of the Ropedia Xperience-10M Task Suite by Chaoyue He.

It is a research adapter over a camera-pose proxy forward-dynamics objective. It is not a JSON action classifier, not a robot policy, not a standalone Cosmos3-Super base model, and not counted as a 20-task matrix method row.

Role In The Project

—Project layer: selected-128 model-diagnostic work.
—Model family: Cosmos3-Super forward dynamics.
—Objective: camera-pose-conditioned future-vision velocity prediction.
—Purpose: test whether Xperience-10M public-safe derived targets can support a world-model-style adapter track alongside the 20-task result matrix.

Run Identity

—Target repo: cy0307/ropedia-cosmos3-super-forward-dynamics-lora-128ep
—Base model: nv-community/Cosmos3-Super
—Dataset run: xperience10m_cosmos3_camera_pose_targets_20260608
—Train run: xperience10m_cosmos3_super_forward_dynamics_lora_128ep_train1epoch_256_attn_full8gpu_20260608
—Eval run: xperience10m_cosmos3_super_forward_dynamics_lora_128ep_train1epoch_256_attn_full8gpu_20260608_eval_test_full_fsdp
—Dataset contract: xperience10m_camera_pose_forward_dynamics_v1
—Objective: camera_pose_conditioned_future_vision_velocity_lora

Data Scope And Metrics

—Train rows: 2,848.
—Validation rows: 512.
—Held-out test rows: 448.
—Held-out test episodes: 14.
—Target kind: slam_camera_pose_delta_proxy_v1.
—Test forward-dynamics MSE: 3.6853.
—Validation forward-dynamics MSE: 4.0082.
—Train final loss: 1.0785.
—Adapter parameters: 26,214,400.

Raw Xperience-10M MP4/HDF5/RRD files and Cosmos3-Super base weights are not included.

Files

—pytorch_lora_weights.safetensors: LoRA adapter tensor state dict.
—training_metadata.json: training run metadata.
—adapter_repair_audit.json: shape repair audit for the FSDP-saved tensors.
—eval_metrics.json and val_metrics.json: held-out loss summaries.
—verified_result_summary.json: public-safe result summary.
—package_audit.json: verified package audit.
—target_manifest.json: camera-pose target manifest.

Interpretation Note

Compare this repo to world-model or forward-prediction branches, not to Qwen3 structured JSON accuracy. The metric is a camera-pose proxy forward-dynamics loss. It is useful as evidence for a future Cosmos-style diagnostic track, but it is not a final full-corpus model-quality claim.

Loading Note

Use the repository scripts that produced the adapter, especially scripts/omni/train_cosmos3_super_forward_dynamics_lora.py and scripts/omni/eval_cosmos3_super_forward_dynamics_lora.py, to map these LoRA tensors onto the staged nv-community/Cosmos3-Super runtime. The tensor file is not a plug-and-play Diffusers pipeline by itself.

Related Repositories

—Project website: https://chaoyue0307.github.io/ropedia-xperience-10m-task-suite/
—GitHub source: https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite
—HF Space: https://huggingface.co/spaces/cy0307/ropedia-xperience-10m-task-suite
—Artifact dataset: https://huggingface.co/datasets/cy0307/ropedia-xperience-10m-task-suite-artifacts
—Baseline model repo: https://huggingface.co/cy0307/ropedia-xperience-10m-task-baselines
—Consolidated weights/results: https://huggingface.co/cy0307/ropedia-xperience-10m-weights-results
—Qwen3-Omni LoRA adapter: https://huggingface.co/cy0307/ropedia-qwen3-omni-lora-128ep
—Cosmos3-Super LoRA adapter: https://huggingface.co/cy0307/ropedia-cosmos3-super-forward-dynamics-lora-128ep
—Upstream gated Xperience-10M dataset: https://huggingface.co/datasets/ropedia-ai/xperience-10m
—Public sample dataset: https://huggingface.co/datasets/ropedia-ai/xperience-10m-sample

Citation

If you use this Cosmos3-Super forward-dynamics adapter, cite it as:

bibtex

@misc{he2026ropedia_cosmos3_super_forward_dynamics_lora,
  author       = {Chaoyue He},
  title        = {Ropedia Xperience-10M Cosmos3-Super Forward-Dynamics LoRA},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/cy0307/ropedia-cosmos3-super-forward-dynamics-lora-128ep},
  note         = {Camera-pose proxy forward-dynamics LoRA artifact for the Ropedia Xperience-10M Task Suite}
}

Also cite the main Ropedia Xperience-10M Task Suite and the upstream Ropedia Xperience-10M dataset according to their citation guidance.