req-metrics 0.1.0__tar.gz

req_metrics-0.1.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Angelos-Nikolaos Kanatas
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

req_metrics-0.1.0/NOTICE

@@ -0,0 +1,26 @@
+req-metrics
+Copyright 2026 Angelos-Nikolaos Kanatas
+
+This product includes software derived from DADApy
+(https://github.com/sissa-data-science/DADApy),
+Copyright 2021-2023 The DADApy Authors, licensed under the Apache License,
+Version 2.0 (http://www.apache.org/licenses/LICENSE-2.0).
+
+The GRIDE likelihood solver and the TwoNN fitting procedure in
+src/req_metrics/metrics/dimension.py are ports of DADApy's
+_utils/utils.py (_filter_mus, _neg_dloglik_did, _argmax_loglik,
+_fisher_info_scaling) and id_estimation.py (_compute_id_2NN,
+return_id_scaling_gride). Changes: rewritten on torch tensors, neighbor
+distances taken from a shared precomputed table instead of recomputed per
+call, scales exposed by their outer neighbor rank, results returned as
+MetricResult.
+
+Method references for every metric are listed in docs/references.bib; the
+toolkit asks users to cite the original paper of each metric they report.
+
+The bias-corrected participation-ratio formulas in req_metrics/metrics/spectral.py
+(participation_ratio with correction="row" or "both") follow the estimators of
+Chun, Canatar, Chung and Lee (ICLR 2026) and their reference implementation
+https://github.com/badooki/dimensionality (MIT License, Copyright (c) 2026 Chan Woo
+Chun). The quartic index sums are computed here in closed form from the Gram matrix
+and column moments.

req_metrics-0.1.0/PKG-INFO

@@ -0,0 +1,288 @@
+Metadata-Version: 2.4
+Name: req-metrics
+Version: 0.1.0
+Summary: Label-free representation-quality metrics for layer-wise analysis of pretrained models and for monitoring during training
+Author: Angelos-Nikolaos Kanatas
+License-Expression: Apache-2.0
+Project-URL: Repository, https://github.com/angeloskanatas/req-metrics
+Project-URL: Documentation, https://github.com/angeloskanatas/req-metrics/tree/main/docs
+Project-URL: Paper, https://arxiv.org/abs/2608.14819
+Project-URL: Layer-wise results, https://angeloskanatas.github.io/music-fms-layer-eval/
+Keywords: representation learning,self-supervised learning,intrinsic dimension,effective rank,layer-wise analysis,training monitoring
+Classifier: Development Status :: 4 - Beta
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Intended Audience :: Science/Research
+Classifier: Operating System :: OS Independent
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: NOTICE
+Requires-Dist: numpy>=1.24
+Requires-Dist: torch>=2.1
+Requires-Dist: scipy>=1.10
+Provides-Extra: lightning
+Requires-Dist: lightning>=2.2; extra == "lightning"
+Provides-Extra: plot
+Requires-Dist: matplotlib>=3.7; extra == "plot"
+Provides-Extra: dev
+Requires-Dist: pytest>=7; extra == "dev"
+Requires-Dist: ruff>=0.16; extra == "dev"
+Requires-Dist: mypy>=1.8; extra == "dev"
+Requires-Dist: lightning>=2.2; extra == "dev"
+Requires-Dist: tensorboard>=2.14; extra == "dev"
+Requires-Dist: pandas>=2.0; extra == "dev"
+Requires-Dist: scikit-learn>=1.3; extra == "dev"
+Dynamic: license-file
+
+# req-metrics
+
+Label-free metrics of representation quality, for analyzing the layers of trained models and for
+monitoring self-supervised training.
+
+[![tests](https://github.com/angeloskanatas/req-metrics/actions/workflows/ci.yml/badge.svg)](https://github.com/angeloskanatas/req-metrics/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/req-metrics.svg)](https://pypi.org/project/req-metrics/)
+[![license](https://img.shields.io/badge/license-Apache--2.0-blue.svg)](https://github.com/angeloskanatas/req-metrics/blob/main/LICENSE)
+[![python](https://img.shields.io/badge/python-3.10%2B-blue.svg)](https://github.com/angeloskanatas/req-metrics/blob/main/pyproject.toml)
+[![arXiv](https://img.shields.io/badge/arXiv-2608.14819-b31b1b.svg)](https://arxiv.org/abs/2608.14819)
+
+When you use a pretrained model as a frozen feature extractor, you have to decide which layer to
+read. When you train a self-supervised model, you want to know whether the run is healthy and which
+checkpoint to keep. Both questions are usually answered by training labeled probes. The metrics in
+this library answer them from the embeddings alone, with a few thousand unlabeled samples.
+
+req-metrics has 39 metrics. They cover the spectrum of the embeddings (effective rank or RankMe,
+α-ReQ, anisotropy), intrinsic dimension (TwoNN, GRIDE, MLE), neighborhoods and clusters, invariance
+and discriminability over augmented views (LiDAR, InfoNCE, DiME), pitch-transposition equivariance
+(PTE), and the similarity between two representations (CKA, SVCCA, information imbalance). Each one
+is a PyTorch function on tensors, so it works with embeddings from any model.
+
+On top of the metrics, `rq.compute` draws a seeded subsample, shares spectra and neighbor tables
+between metrics, and works at three levels: one vector per sample, the tokens of each sample, or the
+tokens of all samples. It returns records that keep the parameters, sample size and citation of every
+value.
+
+The library was written for the paper [What Makes a Good Layer?](https://arxiv.org/abs/2608.14819)
+(ISMIR 2026) and includes more metrics than the paper uses.
+
+## Installation
+
+Requires Python 3.10+ and PyTorch 2.1+.
+
+```bash
+pip install req-metrics
+pip install "req-metrics[lightning]"   # with the Lightning callback
+```
+
+The development version:
+
+```bash
+pip install "req-metrics @ git+https://github.com/angeloskanatas/req-metrics.git"
+```
+
+## Quick start
+
+### Analyzing a trained model
+
+Pass the embeddings of each layer as a dictionary keyed by layer index. The `model` and `pooling`
+labels are stored in the record.
+
+```python
+import req_metrics as rq
+
+layers = {0: z0, 1: z1, 2: z2}                       # (N, D) per layer: one vector per sample
+rec = rq.compute(layers, ["effective_rank", "intrinsic_dimension/gride", "anisotropy/spectral", "self_clustering"],
+                 n_items=10000, seed=42, model="my-encoder", pooling="time-mean")
+rec.profile("effective_rank")                         # [(layer, value), ...]
+print(rec.to_markdown())                              # layers x metrics; also to_csv, to_json, to_pandas
+
+tokens = {0: [t0_s0, t0_s1, ...], 1: [...]}          # (T_i, D) frames or patches per sample
+rq.compute(tokens, ["trajectory_curvature", "effective_rank"], level="sample", n_items=2000)
+rq.compute(tokens, ["effective_rank"], level="population", n_tokens=10000)
+
+views = {0: v0, 1: v1}                                # (q, N, D) augmented views of the same samples
+rq.compute(views, ["lidar", "infonce"], views=rq.ViewSpec(source="shared", augmentations=("PitchShift(-4..4)",), q=10))
+
+rq.compute(shifted_layers, ["pte"], shifts=rq.ShiftSpec("waveform pitch shift", semitones=tuple(range(1, 12))))
+rq.compute_pairs(layers, metrics=["information_imbalance", "cka"])          # every pair of layers
+rq.compute_pairs(audio_layers, text_layers, metrics=["neighborhood_overlap"],  # two models or modalities
+                 params={"neighborhood_overlap": {"k": 10, "l2": True}}, model="audio", model_b="text")
+rq.calibrate(rq.cka, a, b, groups=clip_ids)          # against permutations of whole clips (Gröger et al., arXiv:2602.14486)
+rq.jackknife(rq.information_imbalance, a, b)          # mean and spread over five half-samples (Acevedo et al., arXiv:2505.17101)
+rq.compute(tokens, ["dense_structure"], level="population")  # needs the sample of every token
+
+rq.convergence(z1, "effective_rank").to_markdown()   # the value against the sample size
+rq.top_layers(rec, "intrinsic_dimension/gride", k=3)
+p = rq.protocols.get("kanatas2026")                  # the metrics and parameters of a published analysis
+rq.compute(layers, p.names("sequence"), params=p.params("sequence"), n_items=p.n_items)
+```
+
+Single metrics can also be called directly and return `MetricResult(value, extras)`, for example
+`rq.effective_rank(x)`, `rq.lidar(views)` or `rq.pte(z, shifted)`. The functions in
+`req_metrics.layouts` turn spectrogram-patch grids and frame sequences into points, trajectories or
+token sets.
+
+### Monitoring training
+
+The Lightning callback hooks the blocks of the model and keeps a fixed subset of the training set
+for monitoring. At the start of training and then on a schedule, it computes every metric on every
+layer and logs one scalar per layer and metric (`layer_metrics/<metric>_layer_<l>`). On Weights &
+Biases it also plots the depth profile at each sweep, and with `spectra=256` the singular spectrum of
+every layer.
+
+```python
+from req_metrics.integrations.lightning import LayerMonitorCallback
+
+trainer = pl.Trainer(callbacks=[LayerMonitorCallback(
+    ["effective_rank", "intrinsic_dimension/mlid", "anisotropy/spectral"], layers="backbone.blocks", pool="cls",
+    n_items=5000, every_n_epochs=5, sweep_steps=(100, 300, 1000, 3000, 10000),
+    model_attr="backbone",
+    view_metrics=["lidar"], augment=my_augment, q=10,   # views: the objective's own positives
+    online=True)])                                      # also the training-batch buffer, as a collapse indicator
+```
+
+Without Lightning, `LayerMonitor` does the same with a forward function and a data loader. With
+`drift_metrics`, each layer is also compared with its state at the first (or the previous) sweep, as
+in SVCCA (Raghu et al., NeurIPS 2017):
+
+```python
+mon = rq.LayerMonitor(model.blocks, pool="cls", metrics=["effective_rank", "intrinsic_dimension/mlid"],
+                      n_items=5000, drift_metrics=["cka"])
+mon.sweep(model, monitor_loader, step=epoch, sinks=[rq.wandb_sink(history=mon.history), rq.csv_sink("sweeps.csv")])
+mon.profiles("effective_rank")                       # {step: [(layer, value), ...]}
+mon.profiles("cka")                                  # drift of every layer from its first sweep
+```
+
+A few things to keep in mind when monitoring:
+
+- Spectral and neighbor metrics that center the embeddings miss a collapse onto a single shared
+  vector. Log `normalized_std` (Chen and He, 2021) or `anisotropy/cosine` next to them.
+- Each record also holds `delta_first`, the change since the first sweep, and `drawdown`, the drop
+  below the best value so far. Absolute collapse thresholds vary by dataset (De Melo Costa et al.,
+  2026), so a threshold on the drawdown is easier to set. Log it with `log_extras=("drawdown",)`.
+- A hooked layer must return one row per item. If the trainer flattens the frames of a clip into
+  the batch, pool them back into a (B, T, D) trajectory and use `level="sample"`.
+- Block outputs are taken before the final normalization layer, which can change the spectrum.
+  Compare the last block with the normalized output.
+- Spectral metrics warn when a layer has fewer samples than dimensions, and view metrics warn when
+  two passes return identical views.
+
+The per-layer values are also written to `trainer.callback_metrics`, so `EarlyStopping` and
+`ModelCheckpoint` can use them:
+
+```python
+EarlyStopping(monitor="layer_metrics/normalized_std_layer_12", mode="max", divergence_threshold=0.2 / D**0.5)
+ModelCheckpoint(monitor="layer_metrics/effective_rank_layer_12", mode="max", save_top_k=1)  # only where rank tracks your probe
+```
+
+With `mode="max"`, `divergence_threshold` stops training when the value falls below it. For
+`normalized_std`, 1/sqrt(D) is the value of an embedding that has not collapsed. Layer indices in the
+keys are zero-padded (`layer_07` in a 12-block model). Between sweeps the last value is kept, so
+match the checkpoint schedule to the sweep schedule. A resumed run restores the sweep history and the
+drift reference from the checkpoint.
+
+### Ranking runs and layers
+
+- `rq.rank_runs({name: records}, "effective_rank", layer=12)` orders runs or checkpoints by a
+  metric at one layer. This is the selection rule of RankMe and LiDAR. Use the layer you read
+  downstream.
+- `rq.top_layers` ranks the layers of one model.
+- `rq.top_local_maxima(values, t=3, window=2)` keeps the best local maxima over checkpoints or
+  layers, the rule Dai et al. (NeurIPS 2025) use with `dense_structure`.
+
+The direction is an argument, because whether higher is better depends on the task and on how the
+model was trained. If you have a few hundred labeled examples, check the ranking against a probe
+(`docs/DESIGN.md`, section 4b). Near-equal values should be read as ties.
+
+`pool` can be `cls`, `mean`, `max`, `last`, `tokens`, a grid readout (`gap`, `freq_concat_mean`,
+`partitioned`, `freq_concat`, `freq_mean`, with `pool_kwargs={"grid": (F, T)}`) or any function
+without parameters. `augment` is any function you supply. Results can go to `csv_sink`, `json_sink`,
+`tensorboard_sink`, `wandb_sink` or any function of `(records, step)`. Monitoring and post-hoc
+analysis produce records with the same schema. `docs/DESIGN.md` explains the choices behind the fixed
+subset, the training-batch buffer, the sweep schedule and reading every layer.
+
+## Metrics
+
+| Group | Metrics | Input |
+|---|---|---|
+| Spectral | `alpha_req`, `anisotropy/spectral`, `effective_rank`, `eigenvalue_early_enrichment`, `matrix_entropy`, `participation_ratio` | `(N, D)` points |
+| Intrinsic dimension | `intrinsic_dimension/twonn`, `intrinsic_dimension/gride`, `intrinsic_dimension/mle`, `intrinsic_dimension/mlid`, `intrinsic_dimension/mst` | `(N, D)` points |
+| Relational | `anisotropy/cosine`, `normalized_std`, `self_clustering`, `uniformity` | `(N, D)` points |
+| Clustering | `cluster_quality`, `dense_structure` | `(N, D)` points; `dense_structure` takes the token fields of many samples at the population level |
+| Trajectory | `trajectory_curvature` | `(T, D)` per sample, time-ordered |
+| Views | `alignment`, `dime`, `infonce`, `lidar` | `(q, N, D)` augmented views; `dime` takes q = 2 |
+| Equivariance | `pte` | embeddings of transformed copies (pitch shifts in the paper) |
+| Representation pairs | `cka`, `cycle_knn`, `information_imbalance`, `neighborhood_overlap`, `procrustes`, `ridge_predictivity`, `rsa`, `svcca` | two representations of the same items: layers, checkpoints, models or modalities |
+| Functional | `jacobian_effective_rank` | the model and its inputs, during monitoring |
+| Token fields | `cls_patch_cosine`, `token_cosine`, `token_gram_drift`, `token_norm_outliers` | `(T, D)` token fields per sample; `token_gram_drift` also takes a reference field and is called directly |
+| Distribution | `embedding_norm`, `gaussianity`, `sparsity` | `(N, D)` points |
+
+`rq.list_metrics()`, `rq.get_metric(name)` and `rq.describe(name)` show the input, preprocessing,
+citation and card of each metric. Different estimators of the same quantity share a prefix and are
+named after their method, for example `intrinsic_dimension/twonn` and `intrinsic_dimension/gride`.
+None of them is the default. Options such as a bias correction are arguments, and the record keeps
+the value under every option.
+
+## Documentation
+
+- `docs/METRICS.md`: the definition, protocol and citation of every metric.
+- `docs/metrics/`: one card per metric, generated from the registry.
+- `docs/VIEWS.md`: how augmented views and pitch-shifted copies are passed in.
+- `docs/DESIGN.md`: the design choices, the cost of each metric and the sample sizes.
+- `examples/`: from per-layer embedding files to records, and correlating layer profiles with
+  per-layer downstream scores.
+
+## Reproducing the paper
+
+`rq.protocols.get("kanatas2026")` fixes the metric variants, parameters and sample sizes used in
+the paper. The per-layer results of every model and task are on the
+[project page](https://angeloskanatas.github.io/music-fms-layer-eval/), and `Records.to_atlas_json`
+writes records in the same format.
+
+## Related tools
+
+- [DADApy](https://github.com/sissa-data-science/DADApy) and
+  [scikit-dimension](https://github.com/scikit-learn-contrib/scikit-dimension): intrinsic-dimension estimators.
+- [reptrix](https://github.com/BARL-SSL/reptrix): RankMe, α-ReQ and LiDAR for PyTorch models.
+- [stable-pretraining](https://github.com/galilai-group/stable-pretraining): Lightning callbacks for RankMe
+  and LiDAR on a queue of training batches.
+- [lightly](https://github.com/lightly-ai/lightly): a collapse indicator for self-supervised training.
+- [information_flow](https://github.com/OFSkean/information_flow) (Skean et al., ICML 2025): layer-wise
+  entropy, curvature and InfoNCE for language models.
+- [synesis](https://github.com/chrispla/synesis) (Plachouras et al., IJCNN 2025): probe-based
+  informativeness, equivariance, invariance and disentanglement.
+
+## Development
+
+```bash
+git clone https://github.com/angeloskanatas/req-metrics.git && cd req-metrics
+pip install -e ".[dev]"
+pre-commit install
+ruff check src tests examples scripts && ruff format --check src tests examples scripts
+pytest
+```
+
+`CONTRIBUTING.md` explains how to add a metric.
+
+## Citation
+
+If you use req-metrics, please cite the paper below and the original paper of each metric you
+report. Every metric's citation keys point into `docs/references.bib`, and `CITATION.cff` cites the
+software.
+
+```bibtex
+@inproceedings{kanatas2026goodlayer,
+  title     = {What Makes a Good Layer? Assessing the Layer-Wise Intrinsic Properties of Music Foundation Models},
+  author    = {Kanatas, Angelos-Nikolaos and Kong, Yuexuan and Alonso-Jim{\'e}nez, Pablo and Serra, Xavier and Bogdanov, Dmitry},
+  booktitle = {Proceedings of the International Society for Music Information Retrieval Conference (ISMIR)},
+  year      = {2026},
+  note      = {arXiv:2608.14819}
+}
+```
+
+## License
+
+Apache License 2.0. Code adapted from other projects is credited in `NOTICE`.