4esv commited on
Commit
d1d5a6f
·
verified ·
1 Parent(s): 08dc487

Rameau v1: 21,940 records, 4 configs, verified gold, eval harness

Browse files
Files changed (2) hide show
  1. README.md +11 -10
  2. src/harmony_dataset/export.py +8 -7
README.md CHANGED
@@ -83,11 +83,11 @@ Load one with `load_dataset("4esv/rameau", "<config>")`. Default: `notes_to_rn`.
83
 
84
  | config | task | rows |
85
  |---|---|---|
86
- | `symbol_to_rn` | key + chord symbols -> Roman numerals + cadence (easy: chord quality is given) | 5715 |
87
- | `notes_to_rn` | key + spelled notes -> Roman numerals + cadence (must read each chord) | 5715 |
88
- | `pcset_to_rn` | key + bass-first pitch-class lists -> Roman numerals + cadence (no spelling) | 5715 |
89
- | `key_id` | spelled notes, no key -> identify the key (only key-unambiguous phrases) | 4795 |
90
- | | **total** | **21940** |
91
 
92
  ## Gold labels
93
 
@@ -97,16 +97,17 @@ generates each progression together with its intended analysis. Every chord is
97
  then derived two independent ways with [`music21`](https://web.mit.edu/music21/):
98
  the Roman-numeral figure through the roman engine, and the printed chord symbol
99
  through the chord-symbol parser. An item is kept only if both agree on
100
- pitch-class set and bass. This release: 27480 of
101
- 27480 chords agree. See `VERIFY.md`.
102
 
103
  Built from 845 progression shapes (key-independent), transposed across
104
  keys. All content is synthetic; no third-party corpus is redistributed.
105
 
106
  ## Label convention
107
 
108
- Roman numerals follow the [DCML harmony standard](https://github.com/DCMLab/standards)'s
109
- feature decomposition (`numeral` / `form` / `figbass` / `changes` / `relativeroot`).
 
110
  We follow the notation and copy no DCML data. Major-seventh tonic is `IM7`;
111
  secondary dominants use `/` (e.g. `V7/vi`).
112
  Cadence codes: `PAC` perfect authentic, `IAC` imperfect authentic, `HC` half, `DC` deceptive, `PC` plagal.
@@ -127,7 +128,7 @@ before calling `ChordSymbol` or `Pitch`.
127
  The atomic unit is a shape, a key-independent Roman-numeral sequence. A shape
128
  hashes to exactly one split, so none of its transpositions or task framings
129
  crosses splits. The test split doubles as the benchmark. Rows:
130
- train 14725 / validation 3571 / test 3644.
131
 
132
  ## Known limitations
133
 
 
83
 
84
  | config | task | rows |
85
  |---|---|---|
86
+ | `symbol_to_rn` | key + chord symbols -> Roman numerals + cadence (easy: chord quality is given) | 5,715 |
87
+ | `notes_to_rn` | key + spelled notes -> Roman numerals + cadence (must read each chord) | 5,715 |
88
+ | `pcset_to_rn` | key + bass-first pitch-class lists -> Roman numerals + cadence (no spelling) | 5,715 |
89
+ | `key_id` | spelled notes, no key -> identify the key (only key-unambiguous phrases) | 4,795 |
90
+ | | **total** | **21,940** |
91
 
92
  ## Gold labels
93
 
 
97
  then derived two independent ways with [`music21`](https://web.mit.edu/music21/):
98
  the Roman-numeral figure through the roman engine, and the printed chord symbol
99
  through the chord-symbol parser. An item is kept only if both agree on
100
+ pitch-class set and bass.
101
+ This release: 27,480 of 27,480 chords agree. See `VERIFY.md`.
102
 
103
  Built from 845 progression shapes (key-independent), transposed across
104
  keys. All content is synthetic; no third-party corpus is redistributed.
105
 
106
  ## Label convention
107
 
108
+ Roman numerals follow the feature decomposition of the
109
+ [DCML harmony standard](https://github.com/DCMLab/standards)
110
+ (`numeral` / `form` / `figbass` / `changes` / `relativeroot`).
111
  We follow the notation and copy no DCML data. Major-seventh tonic is `IM7`;
112
  secondary dominants use `/` (e.g. `V7/vi`).
113
  Cadence codes: `PAC` perfect authentic, `IAC` imperfect authentic, `HC` half, `DC` deceptive, `PC` plagal.
 
128
  The atomic unit is a shape, a key-independent Roman-numeral sequence. A shape
129
  hashes to exactly one split, so none of its transpositions or task framings
130
  crosses splits. The test split doubles as the benchmark. Rows:
131
+ train 14,725 / validation 3,571 / test 3,644.
132
 
133
  ## Known limitations
134
 
src/harmony_dataset/export.py CHANGED
@@ -65,7 +65,7 @@ def render_card(res: GenResult, buckets: dict) -> str:
65
  per_split = Counter(r.split for r in res.records)
66
  size_cat = "10K<n<100K" if 10_000 <= total < 100_000 else "1K<n<10K" if total >= 1_000 else "n<1K"
67
  task_rows = "\n".join(
68
- f"| `{t}` | {TASK_BLURB[t]} | {per_task[t]} |" for t in tasks.TASKS
69
  )
70
  return f"""---
71
  pretty_name: "Rameau: Functional Harmony from Notation (Roman Numerals, Cadence, Key)"
@@ -120,7 +120,7 @@ Load one with `load_dataset("4esv/rameau", "<config>")`. Default: `{DEFAULT_CONF
120
  | config | task | rows |
121
  |---|---|---|
122
  {task_rows}
123
- | | **total** | **{total}** |
124
 
125
  ## Gold labels
126
 
@@ -130,16 +130,17 @@ generates each progression together with its intended analysis. Every chord is
130
  then derived two independent ways with [`music21`](https://web.mit.edu/music21/):
131
  the Roman-numeral figure through the roman engine, and the printed chord symbol
132
  through the chord-symbol parser. An item is kept only if both agree on
133
- pitch-class set and bass. This release: {res.attempted_chords - len(res.failures)} of
134
- {res.attempted_chords} chords agree. See `VERIFY.md`.
135
 
136
  Built from {res.shapes} progression shapes (key-independent), transposed across
137
  keys. All content is synthetic; no third-party corpus is redistributed.
138
 
139
  ## Label convention
140
 
141
- Roman numerals follow the [DCML harmony standard](https://github.com/DCMLab/standards)'s
142
- feature decomposition (`numeral` / `form` / `figbass` / `changes` / `relativeroot`).
 
143
  We follow the notation and copy no DCML data. Major-seventh tonic is `IM7`;
144
  secondary dominants use `/` (e.g. `V7/vi`).
145
  Cadence codes: {", ".join(f"`{k}` {v}" for k, v in CADENCE_GLOSS.items())}.
@@ -160,7 +161,7 @@ before calling `ChordSymbol` or `Pitch`.
160
  The atomic unit is a shape, a key-independent Roman-numeral sequence. A shape
161
  hashes to exactly one split, so none of its transpositions or task framings
162
  crosses splits. The test split doubles as the benchmark. Rows:
163
- train {per_split['train']} / validation {per_split['validation']} / test {per_split['test']}.
164
 
165
  ## Known limitations
166
 
 
65
  per_split = Counter(r.split for r in res.records)
66
  size_cat = "10K<n<100K" if 10_000 <= total < 100_000 else "1K<n<10K" if total >= 1_000 else "n<1K"
67
  task_rows = "\n".join(
68
+ f"| `{t}` | {TASK_BLURB[t]} | {per_task[t]:,} |" for t in tasks.TASKS
69
  )
70
  return f"""---
71
  pretty_name: "Rameau: Functional Harmony from Notation (Roman Numerals, Cadence, Key)"
 
120
  | config | task | rows |
121
  |---|---|---|
122
  {task_rows}
123
+ | | **total** | **{total:,}** |
124
 
125
  ## Gold labels
126
 
 
130
  then derived two independent ways with [`music21`](https://web.mit.edu/music21/):
131
  the Roman-numeral figure through the roman engine, and the printed chord symbol
132
  through the chord-symbol parser. An item is kept only if both agree on
133
+ pitch-class set and bass.
134
+ This release: {res.attempted_chords - len(res.failures):,} of {res.attempted_chords:,} chords agree. See `VERIFY.md`.
135
 
136
  Built from {res.shapes} progression shapes (key-independent), transposed across
137
  keys. All content is synthetic; no third-party corpus is redistributed.
138
 
139
  ## Label convention
140
 
141
+ Roman numerals follow the feature decomposition of the
142
+ [DCML harmony standard](https://github.com/DCMLab/standards)
143
+ (`numeral` / `form` / `figbass` / `changes` / `relativeroot`).
144
  We follow the notation and copy no DCML data. Major-seventh tonic is `IM7`;
145
  secondary dominants use `/` (e.g. `V7/vi`).
146
  Cadence codes: {", ".join(f"`{k}` {v}" for k, v in CADENCE_GLOSS.items())}.
 
161
  The atomic unit is a shape, a key-independent Roman-numeral sequence. A shape
162
  hashes to exactly one split, so none of its transpositions or task framings
163
  crosses splits. The test split doubles as the benchmark. Rows:
164
+ train {per_split['train']:,} / validation {per_split['validation']:,} / test {per_split['test']:,}.
165
 
166
  ## Known limitations
167