Upload README.md with huggingface_hub

README.md

@@ -125,4 +125,50 @@ configs:
     path: waveform/train-*
   - split: test
     path: waveform/test-*
+task_categories:
+- image-classification
+license: cc-by-4.0
+tags:
+- bearing-fault-diagnosis
+- vibration
+- signal-to-image
+- cwru
+pretty_name: CWRU Bearing Fault — Perception Representations (signal→VLM)
 ---
+# CWRU — perception representations (visual grounding)
+
+The same CWRU bearing windows rendered as **perception** images — one HF **config** per representation. A frozen VLM vision encoder carries *real but modest* class signal from all of these (shuffle-controlled probe): on a within-condition split it scores high, but under a strict **bearing-wise** split (unseen bearings) it is only slightly above chance — generalizing CWRU fault type across bearings is hard for *every* method, including a purpose-built 1D-CNN. So these are included as **representation-diversity / grounding** data for the foundation model (not as a high-accuracy CWRU classifier). Unlike the `CWRU` (spectrum) repo, they are **not** for compute-then-check CoT (`reasoning` stays empty) — the discriminative signal is non-verbal texture.
+
+## Configs
+```python
+load_dataset("AI4Manufacturing/CWRU-perception", "spectrogram")
+```
+
+| config | records | splits |
+|---|---|---|
+| `spectrogram` | 981 | {'train': 721, 'test': 260} |
+| `scalogram` | 981 | {'train': 721, 'test': 260} |
+| `waveform` | 981 | {'train': 721, 'test': 260} |
+| `reshaped` | 981 | {'train': 721, 'test': 260} |
+
+## Schema (7-field unified record)
+| field | meaning |
+|---|---|
+| `query` | the classification instruction (representation-aware) |
+| `image` | the rendered signal image (bytes embedded) |
+| `annot` | gold fault class: normal / inner_race / outer_race / ball |
+| `reasoning` | chain-of-thought (empty here; filled in the `-annotated` sibling) |
+| `cate` / `task` | `C` / `T-C1` (signal fault classification) |
+| `metadata` | JSON: representation, features, fault_freqs, evidence_tier, sr_nondiagnosable, anomaly, bearing_group, rpm, fs, file_number, window_idx, image_sha256, split |
+
+## Provenance & reproducibility
+Generated **deterministically** by `forge_agent/examples/cwru/convert.py` (`c38f68fad7`) → `forge_model/CWRU/convert_cwru.py` (`ccc0f61ce3`); see `provenance.json` for the full record.
+
+Cleaning (all encoded in the pinned code): excluded the NTN 0.028″ files `3001–3008` (wrong bearing geometry) and the corrupt `191`; flagged Smith & Randall (2015) non-diagnosable (`sr_nondiagnosable`) and acquisition-anomaly files; **bearing-wise leakage-safe split** (all loads + both sample rates of a physical bearing stay on one side). A computed `evidence_tier` (confirmed/weak/absent) cross-checks against the Smith & Randall expert grades.
+
+## Caveats
+- **Ball is the hard class** — Smith & Randall (2015) grade most CWRU ball faults non-diagnosable (intermittent load-zone contact + smeared BSF); those records carry the correct label but a low `evidence_tier`. This is a property of the data, not the converter.
+- **Eval**: use the provided bearing-wise split; the within-condition (per-load) split inflates accuracy.
+
+## Source & license
+Source: **Case Western Reserve University Bearing Data Center** (engineering.case.edu/bearingdatacenter), openly available for research. Diagnosability grades and anomaly flags from W. A. Smith & R. B. Randall, *MSSP* 64–65 (2015) 100–131; leakage-safe split per J. Hendriks et al., *MSSP* 169 (2022) 108732.