DATA_PROVENANCE.md

# Data Provenance — Clover Image Tiny Research Preview

This document records the portable data identity for the exact Stage B
calibration checkpoint in this package. It deliberately contains no local
machine paths, private storage locations, or source-image payloads.

## Clover fine-tuning corpus

| Field | Value |
|---|---|
| Dataset | `Spawning/PD3M` |
| Pinned revision | `2a5eb24a8dccf245acd8e56341761aee06da0bdf` |
| Accepted records | 1,000 image-caption pairs |
| Split | 973 train / 17 validation / 10 test |
| Shards | 1 deterministic checksummed shard |
| Dataset gate license | `CDLA-Permissive-2.0` |
| Item status | CC0-1.0 or Public Domain Mark 1.0, recorded per item |
| Manifest SHA-256 | `50c1249f1cb0d8d690a9acc451ca10c9432eb5a7f4e26f34acb5462096e72322` |
| Preprocessing version | `clover-pd3m-center-crop-512-jpeg95-v1` |

Every accepted manifest row records a stable item ID, caption, split, source
organization and URL, item-level status, source MD5, transformed SHA-256,
original dimensions and MIME type, crop/resize details, filtering result, and
removal status. Preprocessing applies EXIF transpose, a deterministic square
center crop, Lanczos resize to 512×512, and JPEG encoding with stripped
metadata, quality 95, and subsampling 0.

The captions paired with these accepted items were used as the Stage B training
text. They are not represented as a separate, untracked prompt corpus. Neither
the source images nor the gated training shard is distributed in this model
package.

## Checkpoint binding

The data identity above is bound to:

- experiment `clover-kd-20260712T050925Z-01KXABNHP0`;
- optimizer step 500, microstep 4,000, sample position 4,000;
- checkpoint SHA-256
  `4a5b99ff18478742528a0d31c97dcee939b166a51be858721d40ad5984110893`;
- checkpoint-bundle SHA-256
  `384b6515f5f26838aea33ec9a941e06610a20764f0b8637c8b7b0667bfc0d447`;
- resolved-config SHA-256
  `80cf9395d1f587dc0c1d440d9f5b55c55c20703187998509bb306d19d463f597`.

The bundle digest includes checkpoint metadata in addition to the format-2
resume artifacts, preventing the data identity from being relabeled without
changing the approved bundle identity.

## Engineering gallery prompts

The bundled paired contact sheet uses eight fixed project-authored prompts from
the 256-prompt `clover-eval-v1` suite. It is evaluation input, not training
data. Each row fixes the prompt and seed and compares the pinned starting model
on the left with this Stage B checkpoint on the right. The contact sheet
SHA-256 is
`8653d1105b7b0c56e2385a127336dbe9b3a1a03c8518edb4934353a9d8e13bfa`.

This subset checks that the pipeline produces finite, nonblank images. It does
not establish representative quality, prompt alignment, diversity, fairness,
or human preference.

## Foundational upstream provenance limitation

The 1,000-pair manifest fully describes only the additional Clover calibration
data. It does not replace or erase the pretraining lineage already embedded in
the starting student and frozen teacher.

The pinned BK-SDM-Tiny-2M model card names LAION-Aesthetics V2 6.25+ and
2,256,472 pairs, but does not supply the per-record license and provenance
evidence required by Clover's full release policy. Complete per-record
provenance for all foundational data behind the upstream student and teacher is
therefore not available to this project. The upstream checkpoints and their
declared CreativeML OpenRAIL-M licenses are pinned and disclosed, but that is
not equivalent to a complete foundational dataset audit.

This inherited limitation is material. The weights are distributed only as a
**RESEARCH PREVIEW · PRODUCTION PILOT · NOT RELEASE-READY** artifact.