cosmos-reason
Cosmos-Reason1-SFT-Dataset
Dataset Description:
The data format is a pair of video and text annotations. We summarize the data and annotations in Table 4 (SFT), Table 5 (RL), and Table 6 (Benchmark) of the Cosmos-Reason1 paper. We release the annotations for embodied reasoning tasks for BridgeDatav2, RoboVQA, Agibot, HoloAssist, AV, and the videos for the RoboVQA and AV datasets. We additionally release the annotations and videos for the RoboFail dataset for benchmarks. By releasing the dataset, NVIDIA… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Cosmos-Reason1-SFT-Dataset.Cosmos-Reason1-Benchmark
Dataset Description:
The data format is a pair of video and text annotations. We summarize the data and annotations in Table 4 (SFT), Table 5 (RL), and Table 6 (Benchmark) of the Cosmos-Reason1 paper. We release the annotations for embodied reasoning tasks for BridgeDatav2, RoboVQA, Agibot, HoloAssist, AV, and the videos for the RoboVQA and AV datasets. We additionally release the annotations and videos for the RoboFail dataset for benchmarks. By releasing the dataset, NVIDIA… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Cosmos-Reason1-Benchmark.cosmos-reason1-benchmark-mirror
Cosmos-Reason1 Benchmark Mirror
Prepared local mirror for the Cosmos benchmark following the EmbodiedArena layout.
Source input: /home/yons/cosmos-20260413T110714Z-3-001.zip
Contents:
<subtask>/<subtask>_benchmark_qa_pairs.json annotation files
<subtask>/clips/*.mp4 benchmark videos
benchmark.jsonl combined single-file export for local em-eval runs
meta.json validation summary
Subtask summary:
bridgev2: 100 rows, 100 unique videos, 0 duplicated video references
robovqa: 110 rows… See the full description on the dataset page: https://huggingface.co/datasets/thomas-yanxin/cosmos-reason1-benchmark-mirror.Cosmos-Reason1-RL-Dataset
Dataset Description:
The data format is a pair of video and text annotations. We summarize the data and annotations in Table 4 (SFT), Table 5 (RL), and Table 6 (Benchmark) of the Cosmos-Reason1 paper. We release the annotations for embodied reasoning tasks for BridgeDatav2, RoboVQA, Agibot, HoloAssist, AV, and the videos for the RoboVQA and AV datasets. We additionally release the annotations and videos for the RoboFail dataset for benchmarks. By releasing the dataset, NVIDIA… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Cosmos-Reason1-RL-Dataset.vh-features-cosmos-reason2-2b-robometer
vh-features-cosmos-reason2-2b-robometer
Frozen-backbone feature cache for the ego-centric progress value head (RLInf-value-head repo), robot side.
Naming: vh-features-<backbone>-<data>. Sibling with the Qwen backbone on the same rows: vh-features-qwen3vl-2b-robometer.
Backbone: nvidia/Cosmos-Reason2-2B (bf16, frozen). Data: the Robometer processed datasets
(robometer/processed_datasets, arXiv 2603.02115): MetaWorld eval+train, LIBERO-10 + LIBERO-10 failure rollouts,
MIT Franka… See the full description on the dataset page: https://huggingface.co/datasets/bi199797/vh-features-cosmos-reason2-2b-robometer.vh-features-cosmos-reason2-2b-egoverse-v06
vh-features-cosmos-reason2-2b-egoverse-v06
Frozen-backbone feature cache for the ego-centric progress value head (RLInf-value-head repo).
Naming: vh-features-<backbone>-<data>. Siblings: vh-features-qwen3vl-2b-egoverse-v06 (same clips, Qwen backbone),
vh-features-cosmos-reason2-2b-robometer, vh-features-qwen3vl-2b-robometer.
Backbone: nvidia/Cosmos-Reason2-2B (bf16, frozen; Qwen3-VL-2B architecture with NVIDIA's physical-reasoning
post-training). One pooled 2048-d vector per… See the full description on the dataset page: https://huggingface.co/datasets/bi199797/vh-features-cosmos-reason2-2b-egoverse-v06.