Organize model families under versionless paths

.gitattributes

@@ -34,3 +34,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 beat-this-small0/beat-this-small0.pte filter=lfs diff=lfs merge=lfs -text
+beat-this/beat-this-small0.pte filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,83 +1,24 @@
 ---
-license: bsd-2-clause
-library_name: onnxruntime
-pipeline_tag: audio-classification
 tags:
-  - chord-recognition
   - music-information-retrieval
-  - onnx
+  - audio
 ---
 
-# Crema 0.2.0 ONNX for TuneForge
+# TuneForge model catalog
 
-ONNX conversion of the pretrained chord-recognition model distributed in
-[Crema 0.2.0](https://github.com/bmcfee/crema/tree/0.2.0). TuneForge did not
-train this model. This repository supplies a reproducible format conversion,
-fixed inference state, integrity manifest, and validation tooling.
+Versioned, checksummed model releases maintained for
+[TuneForge](https://github.com/grazzolini/tuneforge). TuneForge did not train these models.
 
-## Intended use
+| Family | Task | Format | Runtime | License | Files |
+| --- | --- | --- | --- | --- | --- |
+| Crema 0.2.0 | Chord recognition | ONNX opset 18 | ONNX Runtime 1.29.0 CPU | BSD 2-Clause | [`crema/`](./crema/README.md) |
+| Beat This 1.1.0 `small0` | Beat and downbeat tracking | ExecuTorch PTE FP32 | ExecuTorch 1.4.0 XNNPACK | MIT; runtime BSD 3-Clause | [`beat-this/`](./beat-this/README.md) |
 
-Experimental CPU chord recognition in local TuneForge builds. TensorFlow Crema
-remains TuneForge's default while this conversion receives day-to-day testing.
-The model is not intended for safety-critical or unrelated audio classification.
+Each family directory contains its model card, release manifest, licenses, and
+family-relative `SHA256SUMS`. The Crema release also contains its fixed runtime state. The Beat
+This PTE does not bundle ExecuTorch; application distributions supply the runtime separately.
 
-## Inputs and outputs
-
-- Input: float32 HCQT magnitude, shape `[batch, frames, 216, 2]`, sampled at
-  44,100 Hz with hop length 4,096.
-- Outputs, in order: `chord_tag` (170 classes), `chord_pitch` (12),
-  `chord_root` (13), and `chord_bass` (13).
-- Runtime state: reviewed JSON containing fixed HCQT parameters, the ordered
-  chord vocabulary, and compact transition probabilities. Runtime loading does
-  not use pickle.
-
-## Conversion and validation
-
-The verified Crema wheel and embedded H5 weights are converted with TensorFlow
-2.15.1, tf2onnx 1.16.1, ONNX 1.17.0, and opset 18. ONNX Runtime 1.29.0 uses only
-`CPUExecutionProvider`.
-
-Each release build validates two independent conversions. Required gates:
-
-- reported TensorFlow/ONNX tensor difference for numerical diagnostics;
-- exact argmax parity across all four heads;
-- exact normalized labels, boundaries, extensions, and inversions;
-- exact normalized timelines, which provide identical input to TuneForge's
-  deterministic scorer;
-- maximum aligned confidence difference `0.10`;
-- zero run-to-run timeline drift.
-
-The release manifest records actual artifact hashes and sizes. Rebuilding may
-produce a different ONNX byte hash. Exact argmax and timeline behavior, not
-raw floating-point or byte identity, form the reproducibility gate.
-
-A separate sanitized private checkpoint compared six scorer results directly
-and found zero mismatches. Repository CI does not import or invoke TuneForge's
-scorer.
-
-## Source, authorship, and provenance
-
-- Upstream author: Brian McFee.
-- Crema version: 0.2.0; tag commit
-  `051c91697fd16856a0a1019cc06ee1f11fb52c5f`.
-- Verified wheel SHA-256:
-  `b2787afd0367463438ca2b9b2944c490308eee1f307e5796078ec540a6281484`.
-- Embedded H5 SHA-256:
-  `08b80e5b648e743c89284e9bc0b12b993dad1129157a75e0de70e076b0b8a235`.
-
-Crema's distribution includes the pretrained model but does not provide a
-machine-readable per-track training manifest. Upstream publications and source
-documentation describe evaluation and development using established chord
-annotation corpora; this conversion cannot independently reconstruct or audit
-the complete training set. Do not infer dataset ownership or broader rights
-from this model card.
-
-## License
-
-Crema ships a BSD 2-Clause license covering its distribution, including the
-packaged model files. PyPI metadata labels version 0.2.0 as ISC, conflicting
-with the byte-identical license files in the wheel, source distribution, and
-upstream repository. This release preserves the exact shipped BSD notice in
-`LICENSES/crema-0.2.0-BSD-2-Clause.txt`.
-
-TuneForge's MIT license does not relicense Crema or this converted artifact.
+The Git source, reproducible conversion tooling, and publication workflows live in the
+[TuneForge models repository](https://github.com/grazzolini/tuneforge-models). New model releases
+use maintainer-approved workflows, short-lived OIDC credentials, anonymous byte verification,
+and artifact provenance attestations.

crema/README.md

@@ -0,0 +1,84 @@
+---
+license: bsd-2-clause
+library_name: onnxruntime
+pipeline_tag: audio-classification
+tags:
+  - chord-recognition
+  - music-information-retrieval
+  - onnx
+---
+
+# Crema 0.2.0 ONNX for TuneForge
+
+ONNX conversion of the pretrained chord-recognition model distributed in
+[Crema 0.2.0](https://github.com/bmcfee/crema/tree/0.2.0). TuneForge did not
+train this model. This repository supplies a reproducible format conversion,
+fixed inference state, integrity manifest, and validation tooling.
+
+## Intended use
+
+CPU chord recognition for TuneForge's default Advanced Chords engine. TuneForge packages this
+pinned model and its runtime state whenever Advanced Chords is enabled; the Crema Python package,
+TensorFlow, and Keras are not required at runtime.
+The model is not intended for safety-critical or unrelated audio classification.
+
+## Inputs and outputs
+
+- Input: float32 HCQT magnitude, shape `[batch, frames, 216, 2]`, sampled at
+  44,100 Hz with hop length 4,096.
+- Outputs, in order: `chord_tag` (170 classes), `chord_pitch` (12),
+  `chord_root` (13), and `chord_bass` (13).
+- Runtime state: reviewed JSON containing fixed HCQT parameters, the ordered
+  chord vocabulary, and compact transition probabilities. Runtime loading does
+  not use pickle.
+
+## Conversion and validation
+
+The verified Crema wheel and embedded H5 weights are converted with TensorFlow
+2.15.1, tf2onnx 1.16.1, ONNX 1.17.0, and opset 18. ONNX Runtime 1.29.0 uses only
+`CPUExecutionProvider`.
+
+Each release build validates two independent conversions. Required gates:
+
+- reported TensorFlow/ONNX tensor difference for numerical diagnostics;
+- exact argmax parity across all four heads;
+- exact normalized labels, boundaries, extensions, and inversions;
+- exact normalized timelines, which provide identical input to TuneForge's
+  deterministic scorer;
+- maximum aligned confidence difference `0.10`;
+- zero run-to-run timeline drift.
+
+The release manifest records actual artifact hashes and sizes. Rebuilding may
+produce a different ONNX byte hash. Exact argmax and timeline behavior, not
+raw floating-point or byte identity, form the reproducibility gate.
+
+A separate sanitized private checkpoint compared six scorer results directly
+and found zero mismatches. Repository CI does not import or invoke TuneForge's
+scorer.
+
+## Source, authorship, and provenance
+
+- Upstream author: Brian McFee.
+- Crema version: 0.2.0; tag commit
+  `051c91697fd16856a0a1019cc06ee1f11fb52c5f`.
+- Verified wheel SHA-256:
+  `b2787afd0367463438ca2b9b2944c490308eee1f307e5796078ec540a6281484`.
+- Embedded H5 SHA-256:
+  `08b80e5b648e743c89284e9bc0b12b993dad1129157a75e0de70e076b0b8a235`.
+
+Crema's distribution includes the pretrained model but does not provide a
+machine-readable per-track training manifest. Upstream publications and source
+documentation describe evaluation and development using established chord
+annotation corpora; this conversion cannot independently reconstruct or audit
+the complete training set. Do not infer dataset ownership or broader rights
+from this model card.
+
+## License
+
+Crema ships a BSD 2-Clause license covering its distribution, including the
+packaged model files. PyPI metadata labels version 0.2.0 as ISC, conflicting
+with the byte-identical license files in the wheel, source distribution, and
+upstream repository. This release preserves the exact shipped BSD notice in
+`LICENSES/crema-0.2.0-BSD-2-Clause.txt`.
+
+TuneForge's MIT license does not relicense Crema or this converted artifact.

SHA256SUMS → crema/SHA256SUMS

@@ -1,5 +1,5 @@
+81858cdf4828dc809be65331411ceee9641e896a2fa43a703d331a03410c2dad  LICENSES/crema-0.2.0-BSD-2-Clause.txt
+d2045db926e6574a74dc321165fd0fbaefc6518bc1e137bf25bf51039b557b85  README.md
 a903f9709821fccebb31d4e93d7d783642faaa90859f45f308c0f9131cc7ca59  crema-0.2.0-opset18.onnx
 3744bf9ecb47de7194cb9f250fba26678ea347911af32ec4813645d5e033aca2  crema-0.2.0-runtime-state.json
-3fab560f13ec4a284f62a504676911011323492e1a9b34b3c47059c2972b4640  README.md
-81858cdf4828dc809be65331411ceee9641e896a2fa43a703d331a03410c2dad  LICENSES/crema-0.2.0-BSD-2-Clause.txt
-7f73ccd7da3740c5234b5f3ded22e3a00be42a48cebdc654e1eed1c1b81b66e1  manifest.json
+177a3526ac5714abf8e1db0a1080cd4352f9ec433e1e70d4caecd10f787b1012  manifest.json

manifest.json → crema/manifest.json

@@ -48,8 +48,8 @@
       "size": 1316
     },
     "README.md": {
-      "sha256": "3fab560f13ec4a284f62a504676911011323492e1a9b34b3c47059c2972b4640",
-      "size": 3396
+      "sha256": "d2045db926e6574a74dc321165fd0fbaefc6518bc1e137bf25bf51039b557b85",
+      "size": 3481
     },
     "crema-0.2.0-opset18.onnx": {
       "sha256": "a903f9709821fccebb31d4e93d7d783642faaa90859f45f308c0f9131cc7ca59",