The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
Starling release-set reference tensors
Reference activations that the starlinghq/starling evaluator (model-pipeline evaluate --reference-tensors) compares .NET outputs against for the G2 gate on the accepted P-030 release sets. They are too large for Git and regenerate byte-identically only on the reference CPU, so they are kept here, pinned by commit and SHA-256.
The files are model outputs only. They contain no media: the images and videos themselves are pinned by URL and digest in the repository and fetched from Wikimedia Commons.
Every array is verified before use against a digest already recorded in the repository's protected tree (model-pipeline/protected/). This dataset's commit and each file's SHA-256 are recorded in the hosted-reference.json next to those digests.
| Path | Content | Verified against |
|---|---|---|
embeddinggemma-2/image-fp32-release/tensors/<row>.npz |
image<k>_soft_tokens and token_embeddings (float32) of the 19 rows of the image release set; the members are copied byte for byte from the accepted reference run's .npz (pixel arrays omitted) |
embeddinggemma-2/release/image-fp32-release/tensors.json |
embeddinggemma-2/video-fp32-release/tensors/<row>.npz |
video<k>_soft_tokens and token_embeddings (float32) of the 11 rows of the video release set, copied the same way (frame arrays omitted) |
embeddinggemma-2/release/video-fp32-release/tensors.json |
qwen3.5-0.8b/release-prompts/step0/<id>.f32 |
full-vocabulary step-0 log-softmax (248,320 float32, little-endian) of the 32 release prompts | qwen3.5-0.8b/release/release-prompts/reference/environment.json (step0_full_vectors.sha256) and each case's step0_sha256 in generation.jsonl |
Reference environments: EmbeddingGemma 2 (google/embeddinggemma-2 at 914f7f89142e33e77833254d9c9b90c3cef7303b), sentence-transformers 6.1.0 on PyTorch 2.14.1 CPU fp32, 6 threads. Qwen3.5-0.8B (Qwen/Qwen3.5-0.8B at 2fc06364715b967f1860aea9cf38778875588b17), transformers 5.18.0 on torch 2.14.1+cpu fp32, 6 threads. The full environments are in the protected files above.
Terms: the EmbeddingGemma 2 activations are outputs of a model distributed under the Gemma Terms of Use; the Qwen3.5 activations are outputs of a model released under Apache-2.0. The source media licenses are listed in the release sets' licenses.json.
- Downloads last month
- 93