docs: update README with confidence-scored matching stats and hard-reject methodology

README.md

@@ -4,117 +4,14 @@ tags:
 - music
 - artists
 - metadata
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-dataset_info:
-  features:
-  - name: mb_id
-    dtype: string
-  - name: adb_id
-    dtype: string
-  - name: ma_id
-    dtype: int64
-  - name: progarchives_id
-    dtype: int64
-  - name: jazz_id
-    dtype: string
-  - name: classical_id
-    dtype: int64
-  - name: bandcamp_id
-    dtype: 'null'
-  - name: soundcloud_id
-    dtype: 'null'
-  - name: soundcloud_username
-    dtype: 'null'
-  - name: youtube_channel_id
-    dtype: 'null'
-  - name: sources
-    list: string
-  - name: name
-    dtype: string
-  - name: sort_name
-    dtype: string
-  - name: type
-    dtype: string
-  - name: gender
-    dtype: string
-  - name: disambiguation
-    dtype: string
-  - name: ended
-    dtype: bool
-  - name: aliases
-    list: string
-  - name: country
-    dtype: string
-  - name: area
-    dtype: string
-  - name: begin_date
-    dtype: string
-  - name: end_date
-    dtype: string
-  - name: tags
-    list: string
-  - name: style
-    dtype: string
-  - name: mood
-    dtype: string
-  - name: classical_period
-    dtype: string
-  - name: biography_en
-    dtype: string
-  - name: members
-    dtype: string
-  - name: ipi_codes
-    list: string
-  - name: isni_codes
-    list: string
-  - name: website
-    dtype: string
-  - name: social
-    struct:
-    - name: twitter
-      dtype: string
-    - name: facebook
-      dtype: string
-  - name: images
-    struct:
-    - name: logo_url
-      dtype: string
-    - name: thumb_url
-      dtype: string
-    - name: fanart_url
-      dtype: string
-    - name: banner_url
-      dtype: string
-  - name: urls
-    struct:
-    - name: musicbrainz
-      dtype: string
-    - name: audiodb
-      dtype: string
-    - name: classical
-      dtype: string
-    - name: progarchives
-      dtype: string
-    - name: jazz
-      dtype: string
-    - name: metal_archives
-      dtype: string
-  splits:
-  - name: train
-    num_bytes: 599351174
-    num_examples: 1662320
-  download_size: 294856274
-  dataset_size: 599351174
 ---
 
 # Unified Music Artists
 
 Cross-database artist dataset unifying **MusicBrainz**, **TheAudioDB**, **Metal Archives**, **ProgArchives**, **Jazz**, **Classical Composers**, **Bandcamp**, **SoundCloud**, and **YouTube Music** into one row per artist with flat canonical ID columns.
 
+**1,662,320 total rows** — full outer union across all sources.
+
 ## Canonical ID columns
 
 All nullable — present only when the artist was found in that database:
@@ -134,78 +31,95 @@ All nullable — present only when the artist was found in that database:
 
 ## Coverage
 
-**1,639,519 total rows** — full outer union across all sources. Rows with no MusicBrainz entry are included if present in any other source.
-
-| Source | Rows contributed |
-|--------|-----------------|
-| MusicBrainz | 1,520,428 (primary) |
-| Metal Archives | 139,704 |
-| ProgArchives | 19,367 |
-| Classical Composers DB | 17,159 |
-| SoundCloud | 16,884 |
-| TheAudioDB | 22,387 |
-| YouTube Music | 23,861 |
-| Bandcamp | 11,759 |
-| Jazz DB | 10,941 |
-
-109,190 rows (6.7%) have data from more than one source.
+| Source | Matched to MB | Unmatched (orphan rows) | Total contributed |
+|--------|--------------|------------------------|-------------------|
+| MusicBrainz | 1,520,428 (primary) | — | 1,520,428 |
+| Metal Archives | 16,391 (medium) | +114,751 | 131,142 |
+| Classical Composers DB | 4,135 (medium) | +11,527 | 15,662 |
+| ProgArchives | 4,762 (medium) | +5,681 | 10,443 |
+| Jazz DB | 317 (medium) | +7,642 | 7,959 |
+| TheAudioDB | 10,007 (hard) + 858 (medium) | +2,291 | 13,156 |
 
 ## Methodology
 
 ### Source priority
 
-**MusicBrainz** is the primary source — it has the most complete structured data (aliases, IPI/ISNI codes, begin/end dates, relationships, tags) and is the most authoritative music database. All other sources are joined against MusicBrainz rows first; unmatched rows from each secondary source are appended as orphan rows (full outer union).
+**MusicBrainz** is the primary source — it has the most complete structured data (aliases, IPI/ISNI codes, begin/end dates, relationships, tags). All other sources are joined against MusicBrainz rows first; unmatched rows from each secondary source are appended as orphan rows (full outer union).
 
-### Deduplication and cross-database matching
+### Cross-database matching
 
 #### Hard links — MusicBrainz ID
-**TheAudioDB** stores the MusicBrainz ID directly (`mb_id` field). This is a database-level assertion — when present, it is used as the sole join key with no name comparison.
+
+**TheAudioDB** stores the MusicBrainz UUID directly. When present, it is the sole join key — no name comparison needed. **10,007** artists matched this way.
 
 #### Platform ID mapping via Wikidata SPARQL
-A single Wikidata SPARQL query retrieves all items with a MusicBrainz artist ID (P434) cross-referenced to any of Bandcamp (P3283), SoundCloud (P3040), or YouTube channel ID (P2397). This yields ~53,000 MB→platform mappings in one request, avoiding per-artist API calls across 1.5M rows.
-
-```sparql
-SELECT DISTINCT ?mbid ?bandcamp ?soundcloud ?youtube WHERE {
-  ?item wdt:P434 ?mbid.
-  OPTIONAL { ?item wdt:P3283 ?bandcamp. }
-  OPTIONAL { ?item wdt:P3040 ?soundcloud. }
-  OPTIONAL { ?item wdt:P2397 ?youtube. }
-  FILTER(BOUND(?bandcamp) || BOUND(?soundcloud) || BOUND(?youtube))
-}
-```
 
-Platform profiles (Bandcamp artist pages, SoundCloud user profiles) are then scraped using the mapped slugs/usernames to enrich with images, location, and follower data.
+A single Wikidata SPARQL query retrieves all items with a MusicBrainz artist ID (P434) cross-referenced to Bandcamp (P3283), SoundCloud (P3040), or YouTube channel ID (P2397), yielding ~53,000 MB→platform mappings in one request.
 
-#### Soft matching — norm(name) + country
-For **Metal Archives**, **ProgArchives**, **Jazz**, and **Classical** sources (which do not store MusicBrainz IDs), rows are matched to MusicBrainz entries by:
-1. `norm(name)` exact match — Unicode-normalised, lowercased, punctuation stripped, leading "the " removed
-2. **Plus** same `country` / `area` field
+#### Confidence-scored soft matching
 
-Name alone is used as a fallback only when no other source has matched the same name+country combination, to reduce false positives for common band names.
+For Metal Archives, ProgArchives, Jazz, and Classical sources (which do not store MusicBrainz IDs), rows are matched using a three-tier system:
 
-#### Special-purpose exclusions
-The following MusicBrainz placeholder entries are excluded from the dataset:
-- Names matching `[unknown]`, `[no artist]`, `Various Artists`
-- Any name in bracket notation (`[data]`, `[anonymous]`, `[traditional]`, etc.)
-- Entries with disambiguation containing "special purpose artist" or "language instruction"
+**Medium confidence** (corroborated) — name exact match **plus** at least one of:
+- Same ISO 3166-1 alpha-2 country (all sources normalised to alpha-2 before comparison)
+- `begin_date` / `formed_year` / `birth` within 2 years of MB `begin_date`
+- MB alias list contains the candidate's normalised name
+
+**Rejected** (name matched but hard-rejected by a signal):
+
+| Hard-reject rule | Rationale |
+|-----------------|-----------|
+| Begin/birth year diff > 2 years | Different entity with same name |
+| Candidate end_year predates MB begin_year | Dissolved before the other was founded |
+| Genre-family clash: classical ↔ metal, classical ↔ jazz | Implausible same entity |
+| MB `type=Person`, candidate has `members > 1` | Solo artist matched to a group |
+
+Name-only matches (no corroborating signal) are **never accepted**.
+
+#### Hard-reject counts (name matched, rejected by signals)
+
+| Source | Rejected name-only matches |
+|--------|---------------------------|
+| Metal Archives | 21,079 |
+| ProgArchives | 11,133 |
+| TheAudioDB | 9,637 |
+| Jazz DB | 7,520 |
+| Classical | 4,669 |
+| **Total** | **54,038** |
 
-### Platform ID enrichment
+These 54,038 name collisions would have been false-positive merges under naive name+country matching. Common band names ("Chaos", "Oblivion", "Inferno", "Phoenix") appear dozens of times across countries and eras — year and genre signals correctly separate them.
 
-Beyond Wikidata, Bandcamp and SoundCloud profiles are scraped directly using the mapped slugs to add:
-- **Bandcamp**: `bandcamp_id` (subdomain slug), `location`, `genre`, `tags`, artist image
-- **SoundCloud**: `soundcloud_id` (numeric), `soundcloud_username` (permalink), artist image
+#### Country normalisation
 
-The enrichment pass is idempotent — re-running after scrapers complete additional profiles adds new data without touching already-enriched rows.
+All sources use different country formats. Everything is normalised to ISO 3166-1 alpha-2 before comparison:
 
-## Field normalisation
+- Classical uses alpha-3 (`ITA`→`IT`, `DEU`→`DE`, `GBR`→`GB`, `ENG`→`GB`, `RUS`→`RU`)
+- ProgArchives uses full names (`Italy`→`IT`, `United Kingdom`→`GB`)
+- MusicBrainz and TheAudioDB already use alpha-2
+
+The `country` field in the output is always alpha-2 (or null).
+
+#### Special-purpose exclusions
+
+Excluded from the dataset:
+- `[unknown]`, `[no artist]`, `Various Artists`
+- Any name in bracket notation (`[data]`, `[anonymous]`, `[traditional]`, etc.)
+- Disambiguation containing `"special purpose artist"` or `"language instruction"`
+
+## Field reference
 
 | Field | Notes |
 |-------|-------|
-| `mb_id`, `adb_id`, etc. | Flat top-level columns (not nested `ids` dict) |
 | `sources` | List of database names that contributed to this row |
-| `tags` / `genres` | Lists, deduplicated across sources |
-| `begin_date` / `end_date` | ISO date strings where available |
+| `type` | Person / Group / Orchestra / Choir (from MB) |
+| `country` | ISO 3166-1 alpha-2 |
+| `begin_date` / `end_date` | ISO date strings |
 | `ended` | Boolean — true if the artist/group has dissolved |
+| `tags` | Genre/style tags from MusicBrainz |
+| `genres` / `themes` | Metal Archives genre and lyrical theme strings |
+| `prog_genre` / `jazz_genre` / `classical_period` | Source-specific genre fields |
+| `ipi_codes` / `isni_codes` | Music industry identifiers (MB only) |
+| `aliases` | All known alternate names (MB) |
 
 ## Usage
 
@@ -214,12 +128,15 @@ from datasets import load_dataset
 
 ds = load_dataset("TigreGotico/media-metadata-artists", split="train")
 
-# Artists with Bandcamp pages
-bc = ds.filter(lambda r: r["bandcamp_id"] is not None)
+# Classical composers with MB IDs
+composers = ds.filter(lambda r: r["classical_id"] is not None and r["mb_id"] is not None)
 
 # Metal bands with both MB and MA IDs
 metal = ds.filter(lambda r: r["mb_id"] and r["ma_id"])
 
-# Artists on all three platforms
+# Artists with Bandcamp pages
+bc = ds.filter(lambda r: r["bandcamp_id"] is not None)
+
+# Artists on all three streaming platforms
 triple = ds.filter(lambda r: r["bandcamp_id"] and r["soundcloud_username"] and r["youtube_channel_id"])
 ```