Modelpublic

agentic-learning-ai-lab/jepa-cpp-umaze-parity

sourceHugging Facemitupdated 4d agoView on Hugging Face

0likes414downloads

Model Card

jepa.cpp umaze parity artifacts

Test artifacts for jepa.cpp, an inference runtime for action-conditioned JEPA world models on ggml. jepa.cpp's parity tests and its rollout-latency benchmark (bench/rollout_metric.sh) read these files.

File	What it is
`umaze_dino_channel_straight.v2.f16.gguf`	The world model with F16 weights, GGUF schema v2 (`docs/gguf-schema.md`)
`umaze_dino_channel_straight.v2.f32.gguf`	The same model with F32 weights
`umaze_dino_channel_straight.v2.bf16.gguf`	The same model with BF16 weights
`umaze_ref_gn.npz`	Reference activations from the PyTorch model (`tools/dump_reference.py`), including the CEM and Gauss-Newton planner traces
`umaze_stats.json`	The action and proprio normalization stats the GGUFs embed (`tools/dataset_stats.py`)

The model is a jepa-agent dino_wm world model (dino_channel: DINOv2 ViT-S/14 encoder, channel projector, ViT predictor), trained with a temporal-straightening loss on PointMaze umaze. It is a small, self-trained test model for checking numerical parity. It is not a reference checkpoint for planning results.

SHA-256:

acbef2a9787c861c4a795cbdaa05cacfbb61ee0593d2ccf0b1ffc970f8a4201b  umaze_dino_channel_straight.v2.f16.gguf
6a9934d4e302060dae0cc021fe301bb2a3af0d1dd2c5337c36bc7fd3e2648f1f  umaze_dino_channel_straight.v2.f32.gguf
3d9c750c4b6310871bfb906000120e0550cca81256c85c6a943acdecfc063d29  umaze_dino_channel_straight.v2.bf16.gguf
94dfe3da17902e7088f0e3b9e75db786beb8e25ac793ecabc865a3adce61df5f  umaze_ref_gn.npz
bf07ef1ab499829bb591eb052f003013f779933e2ab86acc34cced2cd72a014d  umaze_stats.json