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

README.md

@@ -164,6 +164,25 @@ Metrics: `exact` (Roman numerals **and** cadence correct — the headline number
 `tonic_acc`, `mode_acc`. The zero-shot prompts are versioned in
 `eval/prompts.py`; parsing rules are documented in `eval/README.md`.
 
+## Baseline results (small-sample probe)
+
+`qwen3:8b` via ollama, thinking disabled, temperature 0, prompt v1, first 60
+test records per config:
+
+| config | exact | chord_acc | cadence_acc |
+|---|---|---|---|
+| `symbol_to_rn` | 0.083 | 0.295 | 0.550 |
+| `notes_to_rn` | 0.000 | 0.116 | 0.117 |
+| `pcset_to_rn` | 0.000 | 0.041 | 0.250 |
+| `key_id` | 0.233 | tonic 0.400 | mode 0.583 |
+
+`llama3.2:3b` on `symbol_to_rn`: exact 0.000, chord_acc 0.137.
+
+The difficulty ladder behaves as designed — chord accuracy falls ~30% -> 12% -> 4%
+as the answer is progressively hidden, and an 8B model cannot reliably produce a
+single fully correct analysis outside the lookup-flavored config. Frontier-model
+results welcome; the harness in `eval/` reproduces these numbers.
+
 ## Reproduce
 
 The full generation pipeline ships in this repo (`src/harmony_dataset/`):

eval/score.py

@@ -50,6 +50,8 @@ def parse_rn(text: str) -> tuple[list[str] | None, str | None]:
     """Extract (labels, cadence) from a model response."""
     text = normalize(text)
     lines = [ln.strip().strip("`") for ln in text.splitlines() if ln.strip()]
+    # drop echoed format placeholders like "<Roman numerals ...>"
+    lines = [ln for ln in lines if not (ln.startswith("<") and ln.endswith(">"))]
     if not lines:
         return None, None
 
@@ -72,6 +74,11 @@ def parse_rn(text: str) -> tuple[list[str] | None, str | None]:
                 if lines[j]:
                     labels_line = lines[j]
                     break
+            else:  # nothing above the cadence line: fall back to below
+                for j in range(cad_idx + 1, len(lines)):
+                    if lines[j]:
+                        labels_line = lines[j]
+                        break
     else:
         labels_line = lines[-1]
 

src/harmony_dataset/export.py

@@ -197,6 +197,25 @@ Metrics: `exact` (Roman numerals **and** cadence correct — the headline number
 `tonic_acc`, `mode_acc`. The zero-shot prompts are versioned in
 `eval/prompts.py`; parsing rules are documented in `eval/README.md`.
 
+## Baseline results (small-sample probe)
+
+`qwen3:8b` via ollama, thinking disabled, temperature 0, prompt v1, first 60
+test records per config:
+
+| config | exact | chord_acc | cadence_acc |
+|---|---|---|---|
+| `symbol_to_rn` | 0.083 | 0.295 | 0.550 |
+| `notes_to_rn` | 0.000 | 0.116 | 0.117 |
+| `pcset_to_rn` | 0.000 | 0.041 | 0.250 |
+| `key_id` | 0.233 | tonic 0.400 | mode 0.583 |
+
+`llama3.2:3b` on `symbol_to_rn`: exact 0.000, chord_acc 0.137.
+
+The difficulty ladder behaves as designed — chord accuracy falls ~30% -> 12% -> 4%
+as the answer is progressively hidden, and an 8B model cannot reliably produce a
+single fully correct analysis outside the lookup-flavored config. Frontier-model
+results welcome; the harness in `eval/` reproduces these numbers.
+
 ## Reproduce
 
 The full generation pipeline ships in this repo (`src/harmony_dataset/`):

tests/test_eval.py

@@ -40,6 +40,13 @@ class TestParseRN:
     def test_garbage(self):
         assert parse_rn("") == (None, None)
 
+    def test_placeholder_echo_is_dropped(self):
+        text = "<Roman numerals separated by single spaces>\ncadence: IAC\nI V65 I"
+        assert parse_rn(text) == (["I", "V65", "I"], "IAC")
+
+    def test_only_placeholder(self):
+        assert parse_rn("<Roman numerals separated by single spaces>") == (None, None)
+
 
 class TestParseKey:
     def test_plain(self):