Upload README.md with huggingface_hub

README.md

@@ -26,19 +26,59 @@ temporal model, no search.
 > playability constraint. CrispASR ships the scorer and leaves the decoder to
 > the caller.
 
-## Attribution (required by CC BY 4.0)
+## Provenance and licence — verifiable, not asserted
+
+**Source artifact** (fetch it and check the hashes yourself):
+
+| | |
+|---|---|
+| file | `best_TabCNN_tablature_trancription_model` (sic — typo is upstream's) |
+| direct URL | <https://zenodo.org/records/11406378/files/best_TabCNN_tablature_trancription_model?download=1> |
+| record | <https://zenodo.org/records/11406378> |
+| size | 3,345,122 bytes |
+| md5 (upstream, from the Zenodo API) | `ce168b2cd426f81a2a78499214e40605` |
+| sha256 (computed on the bytes converted) | `1470a308896629352a811082843eb708cbc2f1aa3092757340055ef76a53ed0c` |
+
+Every GGUF here carries these as metadata (`general.source.url`,
+`general.source.record_url`, `general.source.md5_upstream`,
+`general.source.sha256`), so provenance travels with the artifact.
+
+⚠️ **There is no Zenodo DOI for this record.** `10.5281/zenodo.11406378` looks
+plausible and **404s** — it does not exist. The record's actual DOI is the arXiv
+one, `10.48550/arXiv.2405.14679` (resolves). Cite what resolves.
+
+**Licence evidence.** The Zenodo API reports, for record 11406378:
+
+```json
+"metadata": { "license": {"id": "cc-by-4.0"}, "access_right": "open" }
+```
+
+Zenodo has no per-file licensing (a file object exposes only `checksum`, `id`,
+`key`, `links`, `size`), so the record licence governs every deposited file. And
+the record description states explicitly that the weights are part of the
+deposit:
+
+> "The weights for the best performing model (TabCNN trained with
+> \"GuitarProFX\") in the paper are also provided."
+
+That sentence is also what confirms this is the **GuitarProFX-augmented**
+variant rather than the baseline — the distinction that matters, since the
+baseline collapses from tablature F1 0.748 to 0.447 on real electric guitar
+while the augmented one recovers to 0.585 (DAFx-24).
+
+**Required citation** (the record asks for this explicitly):
+
+> Pedroza HE, Abreu W, Corey R, Roman IR. "Leveraging real electric guitar tones
+> and effects to improve robustness in guitar tablature transcription modeling."
+> In *27th International Conference on Digital Audio Effects (DAFx)*, 2024.
+
+**Upstream chain:**
 
 | | |
 |---|---|
-| **Weights** | from the **EGSet12** dataset record, [doi:10.5281/zenodo.11406378](https://doi.org/10.5281/zenodo.11406378) — CC BY 4.0 |
 | **Model** | TabCNN — Wiggins & Kim, *Guitar Tablature Estimation with a Convolutional Neural Network*, ISMIR 2019 |
 | **Reference implementation** | [`amt-tools`](https://github.com/cwitkowitz/amt-tools) (Cwitkowitz) — MIT |
-| **Augmented training** | Pedroza, Abreu, Corey & Roman, *Leveraging Neural Networks for Guitar Tablature Transcription of Electric Guitar Recordings*, DAFx-24 |
-| **Training corpus** | [GuitarSet](https://doi.org/10.5281/zenodo.3371780) (CC BY 4.0), re-rendered with guitar plugins |
-
-This is the **GuitarProFX-augmented** variant, which is the one worth using: the
-vanilla GuitarSet-trained model collapses from tablature F1 0.748 to **0.447** on
-real electric guitar, while the augmented one recovers to **0.585** (DAFx-24).
+| **Training corpus** | [GuitarSet](https://zenodo.org/records/3371780) — CC BY 4.0 |
 
 ## Files