Add dataset card

README.md

@@ -0,0 +1,119 @@
+---
+license: cc-by-nc-4.0
+pretty_name: jepa.cpp parity fixtures
+size_categories:
+  - n<1K
+tags:
+  - jepa
+  - jepa.cpp
+  - ggml
+  - gguf
+  - parity
+  - golden-references
+  - regression-testing
+---
+
+# jepa.cpp parity fixtures
+
+PyTorch golden reference dumps for [**jepa.cpp**](https://github.com/aselimc/jepa.cpp), a ggml-based
+C/C++ inference engine for the JEPA family. `tests/test-parity` and `tests/test-predictor` replay
+these tensors through the engine and gate per-token cosine, pooled outputs and classifier top-1/top-5
+against per-family thresholds.
+
+Generated from jepa.cpp `main` @ [`4665029`](https://github.com/aselimc/jepa.cpp/commit/4665029) with
+`scripts/dump_reference.py --model all`, in float32 eval mode on 32 CPU threads, no autocast:
+torch 2.13.0+cpu, transformers 5.16.1.
+
+## Contents
+
+| directory | model | samples | size |
+|---|---|---|---|
+| `ref/ijepa-vith14-1k/` | I-JEPA ViT-H/14 | 8 COCO images | 15 MB |
+| `ref/lejepa-vits16/` | LeJEPA ViT-S/16 | 8 COCO images | 7.1 MB |
+| `ref/lewm-pusht/` | LeWorldModel Push-T | 2 images + a 3-frame causal rollout | 3.4 MB |
+| `ref/vjepa2-vitl-fpc64-256/` | V-JEPA 2 ViT-L/16 | 2 clips x {16, 64} frames, encoder + predictor | 350 MB |
+| `ref/vjepa2-vitl-fpc16-256-ssv2/` | V-JEPA 2 ViT-L SSv2 classifier | 2 clips x 16 frames, encoder + pooler + logits | 54 MB |
+| `ref/vjepa2_1-vitb-384/` | V-JEPA 2.1 ViT-B/16 @384 | 1 clip x 16 frames + 2 images | 102 MB |
+| `ref/levjepa-vitl16/` | LeVJEPA ViT-L/16 | 2 clips x 16 frames + 2 still images | 130 MB |
+
+156 files, 660 MB in total.
+
+## Layout
+
+One directory per model, each with a `manifest.json` and one `.npy` per tensor per sample, named
+`<sample>.<tensor>.npy`. All arrays are float32 C-order except `frames_u8` (uint8) and `top5_idx`
+(int64). Shapes carry no batch dimension except `input`, which is stored exactly as fed to the model
+(batch = 1).
+
+The manifest records the model id, the exact hyper-parameters, the preprocessing pipeline that
+produced `input` (resize mode, resampler, crop, mean/std), the PyTorch forward wall time per sample
+(`timing_s.forward_s`, the baseline for the speed tables of the jepa.cpp docs), the frame indices
+sampled from each clip, and the 174 Something-Something-v2 label strings for the classifier. The
+`source` field is the absolute path of the checkout that generated the dump.
+
+Per-model tensor lists and the exact preprocessing recipe:
+[**docs/fixtures.md**](https://aselimc.github.io/jepa.cpp/fixtures/).
+
+## How jepa.cpp consumes it
+
+```bash
+git clone --recursive https://github.com/aselimc/jepa.cpp && cd jepa.cpp
+cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release && cmake --build build -j
+
+scripts/download_fixtures.sh          # this dataset -> tests/fixtures/ref, plus the media it needs
+scripts/download_models.sh small      # GGUFs from https://huggingface.co/jepacpp
+
+cmake -S . -B build && ctest --test-dir build      # parity suites register once ref/ and models/gguf/ exist
+```
+
+A single replay, without ctest:
+
+```bash
+build/test-parity models/gguf/lejepa-vits16-pretrain-in1k-f16.gguf \
+    tests/fixtures/ref/lejepa-vits16 --threads 32 --json out.json
+```
+
+`test-parity` runs two passes per file: first the stored `input` tensor (bypassing preprocessing, so a
+graph bug shows up alone), then jepa.cpp's own preprocessor on the source media (so a preprocessing
+mismatch shows up separately). The second pass needs `tests/fixtures/media/`, which is **not** part of
+this dataset — see below.
+
+## Input provenance
+
+The dumps are model outputs computed on two public research sets. Neither the source images nor the
+source videos are redistributed here; only the reference activations and the decoded frame tensors the
+tests replay are.
+
+| input | source | fetched by |
+|---|---|---|
+| 8 images `coco_*.jpg` | [COCO val2017](https://cocodataset.org/), images subject to their original Flickr terms | tracked in the jepa.cpp git repo |
+| 6 clips (`archery`, `bowling`, `flying_kite`, `high_jump`, `high_jump2`, `marching`) | [`nateraw/kinetics-mini`](https://huggingface.co/datasets/nateraw/kinetics-mini), a small sample of Kinetics-400 | `scripts/download_fixtures.sh` |
+
+The `input` and `frames_u8` arrays inside `ref/` are preprocessed pixels of those 8 images and 6 clips
+(16 or 64 sampled frames each), kept because the parity tests must feed the network exactly the tensor
+PyTorch saw. If you hold rights in any of the underlying material and want it removed, open an issue on
+[github.com/aselimc/jepa.cpp](https://github.com/aselimc/jepa.cpp/issues).
+
+## Licence
+
+`cc-by-nc-4.0`, the most restrictive licence among the checkpoints whose outputs are stored here:
+`ref/ijepa-vith14-1k` and `ref/levjepa-vitl16` are outputs of CC BY-NC 4.0 models
+([facebook/ijepa_vith14_1k](https://huggingface.co/facebook/ijepa_vith14_1k),
+[galilai-group/LeVJEPA-VideoMix-Large](https://huggingface.co/galilai-group/LeVJEPA-VideoMix-Large));
+the other five directories are outputs of MIT / Apache-2.0 models. Research and non-commercial use.
+
+## Regenerating instead of downloading
+
+```bash
+scripts/download_fixtures.sh                      # media
+scripts/download_models.sh all                    # source checkpoints
+.venv/bin/python scripts/dump_reference.py --model all   # ~1 min on 32 cores
+```
+
+Needs the project venv (torch CPU, transformers, av, pillow, timm, einops).
+
+## Links
+
+- Code: <https://github.com/aselimc/jepa.cpp>
+- Documentation: <https://aselimc.github.io/jepa.cpp/>
+- GGUF models: <https://huggingface.co/jepacpp>