a1233646/oracle-coeff-corpus-v1

Oracle Coefficient Corpus V1

Per-sample compressed teacher trajectories of SANA-1600M (1024px, BF16, diffusers), one trajectory per prompt row of a LAION-Aesthetics-derived caption table (479,410 samples, 370,054 distinct normalized prompts). Instead of the full 50-node velocity history, every sample stores the anchor velocity, the endpoint latent, the anchored 50x50 Gram matrix, and weighted-least-squares ("oracle") coefficients of the anchored velocity residual on a frozen, nested K=8 temporal basis (basis/pca_basis.npz, shipped in this repo).

Intended use: training data for coefficient/code prediction heads and trajectory-distillation research. No images are stored, but x_end is a DC-AE latent that decodes to a 1024x1024 generated image.

Layout

manifest.jsonl            one row per sample (479,410 rows), includes the prompt text
manifest_stats.json       corpus-level provenance: teacher settings, hashes, split rule
verify_report.json        pre-collection reproduction gate (run on RTX PRO 6000 before shard generation; it does NOT bitwise-verify the published A100 shards)
basis/pca_basis.npz       frozen temporal basis; sha256 842002964e04909a150e09776603b0e344ad2c7a0208455d0aa72be783fa0097
shards/shard_{i:06d}.safetensors   tensors for manifest rows [i*128, (i+1)*128)
shards/shard_{i:06d}.json          sidecar: sha256 of the .safetensors, provenance, per-sample QC
withheld_shards.txt       indices of shards generated but NOT published in this repo (see "Withheld shards")

The corpus has 3,746 shards (shards 0..3744 hold 128 samples, shard 3745 holds 50; ~146 MiB each, 537 GiB). This repo publishes 3,366 of them (430,770 samples, 481.9 GiB); 380 shards are withheld. manifest.jsonl still lists all 479,410 rows, so rows whose shard index uid // 128 appears in withheld_shards.txt have no tensors here.

Tensor schema (per shard)

Tensor keys are {sample_id}/{name}, where sample_id = r%08d of the manifest row_index.

key	shape	dtype	meaning
grid/t_vel	[50]	F64	velocity query times (shared by all samples in the shard)
grid/dt_actual	[50]	F64	realised BF16 step sizes (shared)
{sid}/a_k8	[8,32,32,32]	F32	oracle WLS coefficients a = Phi_K8^T W R of the anchored residual R(t_j) = v(t_j) - v(t_0); rows :K are the exact solution for any K<8 (nested orthonormal basis)
{sid}/gram	[50,50]	F64	unweighted anchored Gram R R^T over the 50 nodes
{sid}/v0	[32,32,32]	BF16	anchor velocity v(t_0) (CFG-guided output at node 0)
{sid}/x_end	[32,32,32]	BF16	teacher endpoint latent at t=0 (DC-AE f32c32 latent space, before un-scaling)

Reconstruction of the guided velocity at node j with K<=8 basis vectors:

v_hat_j = v0 + np.einsum("k,kchw->chw", basis_K8[j, :K], a_k8[:K])   # basis_K8: [50, 8] from pca_basis.npz

The initial noise x0 is not stored. Regenerate it exactly:

g = torch.Generator(device="cpu").manual_seed(int(latent_seed) % (2**63 - 1))
x0 = torch.randn((1, 32, 32, 32), generator=g, dtype=torch.float32).to(torch.bfloat16)

BF16 tensors cannot be read with the NumPy safetensors backend; use framework="pt".

Sidecar schema (shards/shard_*.json)

status ("complete"), shard_index, sha256 (of the matching .safetensors file), bytes, seconds, frozen_basis_sha256, semantics (field descriptions), provenance (torch/CUDA/GPU, teacher settings, manifest sha256, seed derivation; provenance.model is the generating machine's local checkpoint path), and samples[] with uid, sample_id, row_index, head_split, latent_seed, prompt_sha256, prompt_norm_sha256, in_audit_full_fit, prompt_in_audit_corpus, rel_rms_k4, rel_rms_k8, v0_norm, x_end_norm, seconds_per_sample.

Manifest schema (manifest.jsonl)

uid, sample_id, row_index, prompt, prompt_sha256, prompt_norm_sha256, duplicate_rows_in_tsv, latent_seed, head_split (train 476,780 / val 2,630), in_audit_full_fit, in_audit_full_heldout, in_audit_confirmatory, prompt_in_audit_corpus. uid == row_index; sample uid lives in shard uid // 128. sha256(manifest.jsonl) = ede1cbdc1d3ab2b93ee72133e3dd2daa54516520a5d6c1aade647ae0d98607e1.

head_split is by normalized-prompt group (NFKC, whitespace collapse, casefold), seed 20260902; val excludes every prompt group used to fit or evaluate the basis. The basis itself was fitted on 5,000 of these rows (in_audit_full_fit).

Teacher settings

SANA-1600M 1024px BF16 (diffusers), text encoder Gemma-2-2B-IT (BF16), CFG 4.5 with empty-string unconditional prompt, max_sequence_length=300, no complex human instruction, 50 Euler steps from t=1 to t=0 (ode_step_size=0.02, BF16 time grid), batch size 1. Generated 2026-09-02..07 (local time) with torch 2.10.0+cu128: shards 1..3745 on NVIDIA A100-SXM4-40GB, shard 0 on NVIDIA RTX PRO 6000 Blackwell (see provenance.gpu per sidecar). Re-running on other GPU models or with batch size > 1 does not reproduce bitwise (batching changed coefficients by up to ~15% in the verification run; see verify_report.json).

Loading

The Hub viewer and datasets.load_dataset only read manifest.jsonl (config manifest). Tensor shards are plain files; download them selectively:

hf download a1233646/oracle-coeff-corpus-v1 --repo-type dataset \
  --include "manifest.jsonl" --include "manifest_stats.json" --include "basis/*" \
  --include "shards/shard_00000*" --local-dir ./occ_v1

from huggingface_hub import snapshot_download
from safetensors import safe_open
root = snapshot_download("a1233646/oracle-coeff-corpus-v1", repo_type="dataset",
                         allow_patterns=["manifest.jsonl", "basis/*", "shards/shard_000000.*"])
with safe_open(f"{root}/shards/shard_000000.safetensors", framework="pt") as f:
    a = f.get_tensor("r00000000/a_k8")

Verify each shard against sha256 in its sidecar before use.

Content warning and limitations

• Prompts are web alt-text from a LAION-Aesthetics-derived table, not manually filtered. They include sexual or otherwise sensitive terms, URLs (4.1k rows) and e-mail-like strings (95 rows).
• x_end decodes to images generated from those prompts without a safety filter.
• Coefficients are "oracle" quantities computed from each trajectory's own full velocity history; they say nothing about how predictable they are from the prompt.

Withheld shards

Before upload, every prompt was screened as text (English, multilingual and structural lexicons, then a per-prompt review). 404 rows pair sexual or nudity language with an explicit or possible sign that the subject is a minor, or use exploitation-related terms. Every shard containing such a row is withheld, because shards cannot be split without breaking their recorded sha256: 380 shards (both .safetensors and sidecar .json), listed in withheld_shards.txt. Most rows in those shards are unrelated; they are withheld only because they share a file. The screen read prompt text only and did not inspect generated images, so it is not a guarantee about image content.

Licensing

Not legal advice; each upstream is summarised as the maintainer understands it.

• Tensors were produced with Efficient-Large-Model/Sana_1600M_1024px_BF16_diffusers at revision e2b3c0cbff ("Relicense model under Apache-2.0", 2026-07-31), whose card states Apache License 2.0 with additional reference to the Gemma Terms of Use and Gemma Prohibited Use Policy (text encoder Gemma-2-2B-IT).
• Prompts derive from LAION-Aesthetics metadata (LAION-5B lineage). LAION-5B was withdrawn in December 2023 and re-released as Re-LAION-5B in August 2024. Captions are third-party web text.
• Decoding x_end uses the DC-AE f32c32 SANA autoencoder (Apache-2.0 LICENSE in its repository).

Citation

SANA: arXiv:2410.10629. LAION-5B: arXiv:2210.08402.