# Model and Data Licenses — Clover Image Tiny

This is the license and provenance ledger for the exact public release. The
complete bundled license texts control over this summary. This is not legal
advice.

| Component | Exact identity | License/status | Consequence |
|---|---|---|---|
| Clover Image Tiny weights | Stage B step 500; checkpoint `4a5b99ff18478742528a0d31c97dcee939b166a51be858721d40ad5984110893` | CreativeML OpenRAIL-M derivative | Retain the complete terms, attribution, modification notice, and use restrictions |
| Starting student | `nota-ai/bk-sdm-tiny-2m@aad3e0e8ba61b7cb9f64869dc4e586f8ad9d3665` | Pinned model card declares CreativeML OpenRAIL-M | Derivative-weight obligations apply |
| Frozen teacher | `CompVis/stable-diffusion-v1-4@133a221b8aa7292a167afc5127cb63fb5005638b` | Pinned model card declares CreativeML OpenRAIL-M | Distillation from outputs and activations retains derivative-weight obligations |
| Tokenizer and text encoder | Byte-identical components in the two pinned pipelines | CreativeML OpenRAIL-M lineage | Bundled as upstream model components |
| VAE | Byte-identical component in the two pinned pipelines | CreativeML OpenRAIL-M lineage | Bundled as an upstream model component |
| Clover fine-tuning data | `Spawning/PD3M@2a5eb24a8dccf245acd8e56341761aee06da0bdf`; manifest `50c1249f1cb0d8d690a9acc451ca10c9432eb5a7f4e26f34acb5462096e72322` | Dataset declaration: CDLA-Permissive-2.0; accepted items: CC0-1.0 or Public Domain Mark 1.0 | Dataset and item terms remain separate from the model license; raw data is not bundled |
| Calibration captions | Captions paired with the exact accepted PD3M items | Same recorded item/dataset lineage | Not a separate untracked training corpus |
| Example and packaging code | Files authored by the Clover-Image project | Apache-2.0 | Code license does not relicense model weights or data |
| Runtime dependencies | Pinned packages listed in `requirements.txt` | Their own upstream licenses | Installed separately; no dependency is relicensed by this package |

## Weight-license selection

Both pinned upstream weight repositories declare CreativeML OpenRAIL-M. The
Stage B process modified the starting BK-SDM-Tiny-2M weights using task/output
and feature distillation from the frozen Stable Diffusion v1.4 teacher plus the
licensed calibration corpus. The resulting weights are therefore labeled as a
CreativeML OpenRAIL-M derivative rather than Apache-2.0.

The bundled canonical CreativeML OpenRAIL-M file is 14,385 bytes with SHA-256
`be351ebe7ac01bcdbb018639aadcfd38f136b7dc3f2a3d4d3a24db51d1b210ef`.
The separately bundled Apache-2.0 project-code license is 9,147 bytes with
SHA-256
`e26a0375b714267325c27905ba947fdc41ebc0b7a36940a49ae12379ed6208d6`.

Publishing this artifact as a public release does not create a new weight
license or waive OpenRAIL-M restrictions. Users must review and follow the
complete terms for copying, redistribution, modification, and use.

## Data terms

The fine-tuning dataset gate is `CDLA-Permissive-2.0`, while each accepted item
retains its recorded CC0-1.0 declaration or Public Domain Mark 1.0 status.
Neither the model-weight license nor the project-code license replaces those
data terms. Source images and the gated training shard are not redistributed in
this package.

## Known unresolved provenance scope

The Clover calibration set has an exact 1,000-row manifest. The upstream
foundational data does not have equivalent per-record evidence in this project.
The pinned BK-SDM model card identifies LAION-Aesthetics V2 6.25+ and 2,256,472
pairs, but not the item-level evidence demanded by Clover's full release policy.
The project likewise does not claim a complete per-record audit of every
foundational example behind the upstream teacher.

Pinned model IDs, revisions, and license declarations are evidence of weight
lineage; they are not substitutes for foundational data provenance. This
inherited scope remains explicitly disclosed with the public release.