Add EmoNet Voice Bench results (42/42 classes, pooled r=0.304 rho=0.323) + eval artifacts

README.md

@@ -63,6 +63,9 @@ as [`report/results.csv`](report/results.csv) / [`report/results.json`](report/r
 An interactive report with per-dimension scatter plots is in
 [`report/attrdistill_report.html`](report/attrdistill_report.html).
 
+**External benchmarks are reported [below](#benchmarks)** — these are validation
+numbers on the training distribution, and they do not transfer unchanged.
+
 ---
 
 ## Read this before you use the numbers
@@ -357,6 +360,7 @@ Selected per dimension: `mlp2_h96_64` ×30, `mlp2_h64_32` ×23, `mlp1_h128` ×7,
 | `report/attrdistill_report.html` | interactive report, per-dimension scatter plots |
 | `train_heads.py` | the training script, as run |
 | `provenance/MANIFEST.json` | embedding-matrix manifest: n samples, label column order, source counts, de-dup notes |
+| `bench/emonet/` | EmoNet Voice Bench evaluation: script, per-row predictions, full metrics JSON, label table, data manifest |
 | `provenance/meta_all.parquet` | one row per unique clip: `key`, `tar`, `dim`, `bucket`, `src`, `dur`, `nchar`, `nword` — enough to reproduce the de-duplication and the split |
 
 The 296,422 × 768 embedding matrix and the 296,422 × 61 label matrix are **not**
@@ -395,6 +399,160 @@ roughly the price of one embedding.
 Base encoder: `laion/voiceclap-commercial`. Training data:
 `TTS-AGI/Emotion-Voice-Attribute-Reference-Snippets-DACVAE-Wave`.
 
+
+---
+
+# Benchmarks
+
+Two external benchmarks were run against these heads. Neither is a regression
+benchmark shaped exactly like our training task, so read the methodology before
+the numbers.
+
+## 1. EmoNet Voice Bench — human-expert emotion-intensity regression
+
+**Dataset:** [`t1a5anu-anon/emonet-voice-bench`](https://huggingface.co/datasets/t1a5anu-anon/emonet-voice-bench)
+(public, ungated) — 12,600 items, 12,397 unique clips, 35.8 h. Paper:
+[EmoNet-Voice, arXiv:2506.09827](https://arxiv.org/abs/2506.09827).
+
+**Task, verbatim from the dataset.** Each row carries **exactly one** target emotion
+and an intensity rated `0 = Not Present / 1 = Mildly Present / 2 = Intensely Present`
+by **2–4 of 6 psychology experts**. It is *not* a 40-way argmax and *not* multi-label.
+The paper maps the averaged human intensity to a 0–10 scale (0 → 0, 1 → 5, 2 → 10) and
+reports **MAE, RMSE, Pearson r and Spearman ρ**.
+
+**Our protocol.** For each row we take the *single* head named by that row's emotion
+and score the clip with it — no argmax, no head selection. **42 of 42 benchmark classes
+map 1:1 onto a head**, with only cosmetic renames
+(`Astonishment→Astonishment_Surprise`, `Fatigue→Fatigue_Exhaustion`,
+`Hope→Hope_Enthusiasm_Optimism`, `Intoxication→Intoxication_Altered_States_of_Consciousness`,
+`Malevolence→Malevolence_Malice`, `Pleasure→Pleasure_Ecstasy`,
+`Thankfulness→Thankfulness_Gratitude`, `Emotional Numbness`, `Jealousy & Envy`,
+`Sexual Lust`, `Impatience and Irritability`). The remaining 19 heads are unused.
+All 12,600 rows were scored; **refusal rate 0 %**.
+
+### Results
+
+| aggregation | Pearson r | Spearman ρ |
+|---|--:|--:|
+| **pooled over all 12,600 rows** | **0.304** | **0.323** |
+| macro-averaged over the 42 emotions | 0.386 | 0.374 |
+| macro, restricted to the 2,912 rows with unanimous annotators | 0.548 | 0.488 |
+
+Published baselines (paper Table 5), for orientation:
+
+| model | Spearman ρ | Pearson r | MAE | RMSE | refusal |
+|---|--:|--:|--:|--:|--:|
+| EmpathicInsight-Voice Large | 0.415 | 0.421 | 2.995 | 3.756 | 0 % |
+| EmpathicInsight-Voice Small | 0.418 | 0.414 | 2.997 | 3.757 | 0 % |
+| Gemini 2.5 Pro | 0.417 | 0.416 | 3.008 | 3.785 | 0 % |
+| Gemini 2.0 Flash | 0.355 | 0.350 | 3.608 | 4.453 | 0.01 % |
+| GPT-4o Audio Preview | 0.337 | 0.336 | 3.432 | 4.247 | 27.6 % |
+| GPT-4o Mini Audio Preview | 0.326 | 0.327 | 3.320 | 4.124 | 2.3 % |
+| Hume Voice | 0.274 | 0.231 | 4.744 | 5.474 | 39.2 % |
+| **these heads (pooled)** | **0.323** | **0.304** | see below | see below | **0 %** |
+
+**Honest reading:** on the fair, scale-free comparison these 61-head probes land
+**below Gemini 2.0 Flash and the EmpathicInsight-Voice models, above Hume Voice, and
+roughly level with GPT-4o Mini Audio Preview** — while being a 50 k-parameter MLP on a
+frozen embedding rather than a large audio LLM. The paper does not state whether its
+Table 5 aggregates pooled or macro-over-emotions; we give both, and we quote our
+**pooled** figure in the table above because it is the less flattering of the two.
+
+### MAE / RMSE need calibration and are therefore *not* comparable
+
+Our heads emit the Empathic-Insight 0–7 scale, not 0–10, and the benchmark's rows are
+heavily emotion-present (gold mean 5.26 on the 0–10 scale) while our raw predictions
+average 1.13. Raw MAE is dominated by that scale mismatch, not by ranking quality:
+
+| variant | MAE | RMSE | fair to compare? |
+|---|--:|--:|---|
+| fixed rescale `clip(score/7×10, 0, 10)`, no fitting | 3.884 | 4.587 | yes, but it is mostly measuring the scale gap |
+| one global affine, 2-fold cross-fitted on the benchmark | 2.289 | 2.786 | **no** — uses benchmark labels |
+| per-emotion affine, 2-fold cross-fitted on the benchmark | 2.142 | 2.607 | **no** — uses benchmark labels |
+| per-emotion affine fit on all data (oracle) | 2.124 | 2.579 | **no** — upper bound only |
+
+The calibrated rows would "beat" every model in Table 5. **They are not a win.** The
+LLM baselines are zero-shot and receive no in-domain calibration; we do. We report them
+only to show that most of the raw MAE gap is an offset/scale artefact rather than a
+ranking failure. **Use the correlation table for comparisons.** This is caveat 2 (the
+bucket-balanced bias) showing up in a second domain, in the opposite direction: our
+absolute level is far too *low* here because the benchmark over-samples present-emotion
+clips relative to our training marginal.
+
+### Presence detection (a secondary, non-official metric)
+
+Reframing as "is this emotion present at all?" — clips all annotators rated 0 vs clips
+all annotators rated > 0 — gives a **macro ROC-AUC of 0.810** over the 38 emotions with
+≥ 10 clips on both sides. This is not the benchmark's defined metric; it is reported
+because it is the shape most downstream users actually want (filtering a corpus), and
+because it is threshold-free.
+
+### Contamination check
+
+Identifier spaces are **disjoint**: benchmark clips are named `<8-hex>_enhanced*.mp3`,
+our training keys are `cond_podcastt_*` / `emo_<dim>_b<bucket>_batch*_chunk*`. Zero
+overlap on exact stems and zero overlap on 8-hex tokens. That is evidence of separate
+provenance, **not** proof of zero audio overlap — no audio-fingerprint dedup was run.
+
+**A lineage caveat that is more important than the filename check:** our training
+labels are *Empathic Insight* model annotations, and the EmpathicInsight-Voice models
+were themselves trained on EmoNet-Voice data. So our teacher may have seen clips from
+the same parent corpus as this benchmark. Treat these numbers as *distillation
+transfer measured against human experts*, not as a clean held-out generalisation test
+against an unrelated corpus. The gap between our 0.30 and EmpathicInsight-Voice's 0.42
+is roughly the cost of distilling that annotator into a 50 k-parameter probe.
+
+### Per-emotion results
+
+`presence AUROC` is blank where one side had < 10 clips.
+
+| emotion | n | Pearson r | Spearman ρ | r (unanimous) | presence AUROC |
+|---|--:|--:|--:|--:|--:|
+| `Teasing` | 300 | 0.652 | 0.643 | 0.762 | 0.938 |
+| `Impatience_and_Irritability` | 300 | 0.651 | 0.640 | 0.806 | 0.940 |
+| `Malevolence_Malice` | 300 | 0.582 | 0.559 | 0.787 | 0.922 |
+| `Authenticity` | 300 | 0.577 | 0.464 | 0.819 | 0.939 |
+| `Anger` | 300 | 0.567 | 0.572 | 0.720 | 0.876 |
+| `Amusement` | 300 | 0.535 | 0.508 | 0.799 | 0.903 |
+| `Pain` | 300 | 0.499 | 0.478 | 0.676 | 0.826 |
+| `Distress` | 300 | 0.495 | 0.497 | 0.613 | 0.829 |
+| `Sexual_Lust` | 300 | 0.485 | 0.457 | 0.596 | 0.847 |
+| `Contempt` | 300 | 0.461 | 0.453 | 0.666 | 0.895 |
+| `Disgust` | 300 | 0.457 | 0.460 | 0.701 | 0.865 |
+| `Helplessness` | 300 | 0.454 | 0.438 | 0.631 | 0.856 |
+| `Sadness` | 300 | 0.440 | 0.383 | 0.650 | 0.820 |
+| `Elation` | 300 | 0.438 | 0.446 | 0.605 | 0.820 |
+| `Fear` | 300 | 0.413 | 0.402 | 0.675 | 0.836 |
+| `Jealousy_&_Envy` | 300 | 0.406 | 0.381 | 0.556 | 0.791 |
+| `Fatigue_Exhaustion` | 300 | 0.397 | 0.428 | 0.586 | 0.836 |
+| `Pride` | 300 | 0.396 | 0.368 | 0.625 | 0.811 |
+| `Embarrassment` | 300 | 0.380 | 0.379 | 0.407 | 0.752 |
+| `Arousal` | 300 | 0.370 | 0.339 | 0.552 | 0.823 |
+| `Astonishment_Surprise` | 300 | 0.368 | 0.357 | 0.488 | 0.744 |
+| `Confusion` | 300 | 0.367 | 0.364 | 0.623 | 0.869 |
+| `Bitterness` | 300 | 0.360 | 0.330 | 0.514 | — |
+| `Infatuation` | 300 | 0.358 | 0.377 | 0.367 | 0.782 |
+| `Pleasure_Ecstasy` | 300 | 0.355 | 0.344 | 0.552 | 0.745 |
+| `Longing` | 300 | 0.351 | 0.363 | 0.544 | 0.820 |
+| `Affection` | 300 | 0.342 | 0.287 | 0.553 | 0.789 |
+| `Intoxication_Altered_States_of_Consciousness` | 300 | 0.342 | 0.357 | 0.446 | 0.795 |
+| `Sourness` | 300 | 0.340 | 0.350 | 0.539 | — |
+| `Shame` | 300 | 0.335 | 0.302 | 0.507 | 0.792 |
+| `Doubt` | 300 | 0.315 | 0.324 | 0.382 | 0.735 |
+| `Disappointment` | 300 | 0.309 | 0.289 | 0.465 | 0.763 |
+| `Thankfulness_Gratitude` | 300 | 0.309 | 0.250 | 0.505 | 0.721 |
+| `Interest` | 300 | 0.289 | 0.261 | 0.485 | 0.843 |
+| `Contentment` | 300 | 0.284 | 0.303 | 0.452 | 0.786 |
+| `Concentration` | 300 | 0.277 | 0.276 | 0.334 | — |
+| `Relief` | 300 | 0.277 | 0.263 | 0.452 | 0.735 |
+| `Triumph` | 300 | 0.274 | 0.290 | 0.441 | 0.760 |
+| `Contemplation` | 300 | 0.262 | 0.238 | 0.547 | — |
+| `Awe` | 300 | 0.177 | 0.153 | 0.281 | 0.663 |
+| `Emotional_Numbness` | 300 | 0.171 | 0.217 | 0.139 | 0.754 |
+| `Hope_Enthusiasm_Optimism` | 300 | 0.108 | 0.134 | 0.151 | 0.569 |
+
+Run with [`bench/emonet/eval_emonet.py`](https://huggingface.co/laion/voiceclap-commercial-attribute-heads/blob/main/bench/emonet/eval_emonet.py); raw per-row predictions and the full metrics JSON are in [`bench/emonet/`](bench/emonet).
+
 ---
 
 ## Full results (all 61 dimensions)

bench/emonet/DATA_MANIFEST.json

@@ -0,0 +1,326 @@
+{
+  "dataset": "t1a5anu-anon/emonet-voice-bench",
+  "revision": "c0067f7b2c1b303d77394b42c09836f7be84932c",
+  "license": "cc-by-4.0",
+  "gated": false,
+  "private": false,
+  "downloaded_to": "/e/data1/datasets/playground/mmlaion/schuhmann1/dramabox/headspub/bench/emonet/data",
+  "download_complete": true,
+  "total_bytes": 1077182855,
+  "paper": "arXiv:2506.09827 - EmoNet-Voice: A Fine-Grained, Expert-Verified Benchmark for Speech Emotion Detection",
+  "splits": {
+    "train": {
+      "n_rows": 12600,
+      "note": "ONLY split in repo; it IS the eval/benchmark set despite being named 'train'"
+    }
+  },
+  "configs": [
+    "default"
+  ],
+  "files": [
+    {
+      "path": "data/data/train-00000-of-00003.parquet",
+      "bytes": 366186448,
+      "num_rows": 4200
+    },
+    {
+      "path": "data/data/train-00001-of-00003.parquet",
+      "bytes": 354533930,
+      "num_rows": 4200
+    },
+    {
+      "path": "data/data/train-00002-of-00003.parquet",
+      "bytes": 356462477,
+      "num_rows": 4200
+    }
+  ],
+  "schema": {
+    "audioId": "struct<bytes: binary, path: string> (HF Audio feature; MP3 bytes embedded in parquet)",
+    "label": "string holding a Python-repr list of per-annotator dicts, e.g. \"[{'human-1': {'Shame': 1}}, {'human-2': {'Shame': 2}}]\"; emotion names are URL-encoded (%20, %26)"
+  },
+  "audio": {
+    "encoding": "MP3 (MPEG Layer III), ID3v2 tagged",
+    "storage": "embedded in parquet as bytes; NOT extracted to loose files",
+    "sample_rate_hz": 44100,
+    "channels": 1,
+    "duration_sec": {
+      "min": 0.20800453514739228,
+      "max": 224.35,
+      "mean": 10.235554517150774,
+      "median": 6.348990929705216,
+      "total_hours": 35.82444081002771
+    }
+  },
+  "label_space": {
+    "n_classes": 42,
+    "classes": [
+      "Affection",
+      "Amusement",
+      "Anger",
+      "Arousal",
+      "Astonishment",
+      "Authenticity",
+      "Awe",
+      "Bitterness",
+      "Concentration",
+      "Confusion",
+      "Contemplation",
+      "Contempt",
+      "Contentment",
+      "Disappointment",
+      "Disgust",
+      "Distress",
+      "Doubt",
+      "Elation",
+      "Embarrassment",
+      "Emotional Numbness",
+      "Fatigue",
+      "Fear",
+      "Helplessness",
+      "Hope",
+      "Impatience and Irritability",
+      "Infatuation",
+      "Interest",
+      "Intoxication",
+      "Jealousy & Envy",
+      "Longing",
+      "Malevolence",
+      "Pain",
+      "Pleasure",
+      "Pride",
+      "Relief",
+      "Sadness",
+      "Sexual Lust",
+      "Shame",
+      "Sourness",
+      "Teasing",
+      "Thankfulness",
+      "Triumph"
+    ],
+    "task": "per-(clip, target-emotion) INTENSITY rating. Exactly ONE target emotion per row, rated 0/1/2 by 2-4 experts. NOT multi-label, NOT a 42-way argmax task.",
+    "intensity_scale": {
+      "0": "Not Present",
+      "1": "Mildly Present",
+      "2": "Intensely Present"
+    },
+    "official_gold_mapping": "0->0, 1->5, 2->10 on a 0-10 scale; averaged across annotators (column gold_0_10)",
+    "n_annotators_per_clip": {
+      "2": 4538,
+      "3": 7719,
+      "4": 343
+    },
+    "annotator_ids": [
+      "human-1",
+      "human-2",
+      "human-3",
+      "human-4",
+      "human-5",
+      "human-6"
+    ],
+    "rows_per_class": {
+      "Affection": 300,
+      "Amusement": 300,
+      "Anger": 300,
+      "Arousal": 300,
+      "Astonishment": 300,
+      "Authenticity": 300,
+      "Awe": 300,
+      "Bitterness": 300,
+      "Concentration": 300,
+      "Confusion": 300,
+      "Contemplation": 300,
+      "Contempt": 300,
+      "Contentment": 300,
+      "Disappointment": 300,
+      "Disgust": 300,
+      "Distress": 300,
+      "Doubt": 300,
+      "Elation": 300,
+      "Embarrassment": 300,
+      "Emotional Numbness": 300,
+      "Fatigue": 300,
+      "Fear": 300,
+      "Helplessness": 300,
+      "Hope": 300,
+      "Impatience and Irritability": 300,
+      "Infatuation": 300,
+      "Interest": 300,
+      "Intoxication": 300,
+      "Jealousy & Envy": 300,
+      "Longing": 300,
+      "Malevolence": 300,
+      "Pain": 300,
+      "Pleasure": 300,
+      "Pride": 300,
+      "Relief": 300,
+      "Sadness": 300,
+      "Sexual Lust": 300,
+      "Shame": 300,
+      "Sourness": 300,
+      "Teasing": 300,
+      "Thankfulness": 300,
+      "Triumph": 300
+    },
+    "mean_intensity_per_class": {
+      "Shame": 1.1183,
+      "Helplessness": 1.0269,
+      "Astonishment": 1.0606,
+      "Bitterness": 1.2369,
+      "Concentration": 1.4022,
+      "Contentment": 1.0994,
+      "Disgust": 1.0589,
+      "Confusion": 1.1197,
+      "Affection": 1.1239,
+      "Infatuation": 1.2125,
+      "Thankfulness": 1.0492,
+      "Triumph": 1.1783,
+      "Pride": 1.0856,
+      "Longing": 0.9539,
+      "Sadness": 0.8925,
+      "Anger": 1.2011,
+      "Jealousy & Envy": 1.1639,
+      "Arousal": 0.7769,
+      "Intoxication": 0.92,
+      "Contempt": 1.1728,
+      "Impatience and Irritability": 1.1036,
+      "Emotional Numbness": 0.6581,
+      "Distress": 1.1619,
+      "Teasing": 0.9997,
+      "Pleasure": 0.9189,
+      "Sourness": 1.1617,
+      "Authenticity": 1.0475,
+      "Disappointment": 1.175,
+      "Fatigue": 0.9125,
+      "Interest": 1.0983,
+      "Doubt": 1.02,
+      "Elation": 0.9908,
+      "Sexual Lust": 0.8378,
+      "Fear": 0.9011,
+      "Malevolence": 1.1603,
+      "Amusement": 1.0042,
+      "Contemplation": 1.1425,
+      "Embarrassment": 1.1661,
+      "Awe": 0.8586,
+      "Pain": 1.0086,
+      "Hope": 0.9064,
+      "Relief": 1.0814
+    },
+    "annotation_counts_by_intensity": {
+      "0": 10828,
+      "1": 9009,
+      "2": 13768
+    }
+  },
+  "unique_audio_files": 12397,
+  "duplicate_audio_note": "197 mp3 filenames each appear in 2 rows, always with a DIFFERENT target emotion -> the unit of evaluation is (audio_path,label); clip_id is unique per row",
+  "consensus_subsets": {
+    "unanimous_rows": 2912,
+    "present_all_annotators": 4692,
+    "present_any_annotator": 11475,
+    "absent_all": 1125
+  },
+  "official_metric": {
+    "primary": [
+      "MAE",
+      "RMSE",
+      "Pearson r",
+      "Spearman rho"
+    ],
+    "evidence": "arXiv:2506.09827 (https://arxiv.org/html/2506.09827v1), evaluation section / Table 5: 'our primary evaluation employs metrics suited for regression and correlation analysis on a common scale. The 3-level intensity human judgments (0: Not Present, 1: Mildly Present, 2: Intensely Present) are mapped to a 0-10 scale for this evaluation, becoming 0, 5, and 10, respectively.'",
+    "note": "Primary protocol is regression/correlation, NOT argmax classification. Accuracy figures appear only anecdotally in the abstract."
+  },
+  "labels_parquet": {
+    "path": "/e/data1/datasets/playground/mmlaion/schuhmann1/dramabox/headspub/bench/emonet/labels.parquet",
+    "n_rows": 12600,
+    "columns": [
+      "clip_id",
+      "audio_path",
+      "parquet_file",
+      "row_index",
+      "rel_path",
+      "label",
+      "label_raw",
+      "head_name",
+      "n_annotators",
+      "ratings",
+      "intensity_mean",
+      "intensity_median",
+      "intensity_majority",
+      "intensity_min",
+      "intensity_max",
+      "gold_0_10",
+      "unanimous",
+      "present_any",
+      "present_all",
+      "duration_sec",
+      "sample_rate",
+      "audio_bytes"
+    ],
+    "how_to_get_audio": "open <parquet_file> with pyarrow, take row <row_index> of column audioId -> {'bytes': mp3, 'path': name}. rel_path encodes 'file#row_index'."
+  },
+  "head_mapping": {
+    "Affection": "Affection",
+    "Amusement": "Amusement",
+    "Anger": "Anger",
+    "Arousal": "Arousal",
+    "Astonishment": "Astonishment_Surprise",
+    "Authenticity": "Authenticity",
+    "Awe": "Awe",
+    "Bitterness": "Bitterness",
+    "Concentration": "Concentration",
+    "Confusion": "Confusion",
+    "Contemplation": "Contemplation",
+    "Contempt": "Contempt",
+    "Contentment": "Contentment",
+    "Disappointment": "Disappointment",
+    "Disgust": "Disgust",
+    "Distress": "Distress",
+    "Doubt": "Doubt",
+    "Elation": "Elation",
+    "Embarrassment": "Embarrassment",
+    "Emotional Numbness": "Emotional_Numbness",
+    "Fatigue": "Fatigue_Exhaustion",
+    "Fear": "Fear",
+    "Helplessness": "Helplessness",
+    "Hope": "Hope_Enthusiasm_Optimism",
+    "Impatience and Irritability": "Impatience_and_Irritability",
+    "Infatuation": "Infatuation",
+    "Interest": "Interest",
+    "Intoxication": "Intoxication_Altered_States_of_Consciousness",
+    "Jealousy & Envy": "Jealousy_&_Envy",
+    "Longing": "Longing",
+    "Malevolence": "Malevolence_Malice",
+    "Pain": "Pain",
+    "Pleasure": "Pleasure_Ecstasy",
+    "Pride": "Pride",
+    "Relief": "Relief",
+    "Sadness": "Sadness",
+    "Sexual Lust": "Sexual_Lust",
+    "Shame": "Shame",
+    "Sourness": "Sourness",
+    "Teasing": "Teasing",
+    "Thankfulness": "Thankfulness_Gratitude",
+    "Triumph": "Triumph"
+  },
+  "heads_unused": [
+    "Age",
+    "Background_Noise",
+    "Confident_vs._Hesitant",
+    "Gender",
+    "High-Pitched_vs._Low-Pitched",
+    "Monotone_vs._Expressive",
+    "Recording_Quality",
+    "Serious_vs._Humorous",
+    "Soft_vs._Harsh",
+    "Submissive_vs._Dominant",
+    "Valence",
+    "Vulnerable_vs._Emotionally_Detached",
+    "Warm_vs._Cold",
+    "duration",
+    "score_background_quality",
+    "score_content_enjoyment",
+    "score_overall_quality",
+    "score_speech_quality",
+    "talking_speed"
+  ]
+}

bench/emonet/emonet_labels.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ebcad858a71b7f683cefda0958cf6d25e31cde64dcc6f6be89d9c14e9b53345
+size 317318

bench/emonet/emonet_metrics.json

@@ -0,0 +1,1154 @@
+{
+ "all": {
+  "per_class": {
+   "Affection": {
+    "n": 300,
+    "r": 0.3423618110016751,
+    "rho": 0.28748763875154165,
+    "mae_fixed": 3.7467981098190184,
+    "rmse_fixed": 4.40622252770672,
+    "mae_oracle": 2.088718460385177,
+    "rmse_oracle": 2.5410422028650927,
+    "pred_mean": 1.4165073360999425,
+    "gold_mean": 5.619444444444445
+   },
+   "Amusement": {
+    "n": 300,
+    "r": 0.5347614608335096,
+    "rho": 0.5084536494849943,
+    "mae_fixed": 3.7450283745924633,
+    "rmse_fixed": 4.375549230776509,
+    "mae_oracle": 1.9915500448118981,
+    "rmse_oracle": 2.369039354017009,
+    "pred_mean": 0.9449673793713251,
+    "gold_mean": 5.020833333333333
+   },
+   "Anger": {
+    "n": 300,
+    "r": 0.5665113624595921,
+    "rho": 0.5724392493249925,
+    "mae_fixed": 3.95999262240671,
+    "rmse_fixed": 4.529758285747403,
+    "mae_oracle": 2.0414323543042565,
+    "rmse_oracle": 2.509611240223811,
+    "pred_mean": 1.6009647358953953,
+    "gold_mean": 6.0055555555555555
+   },
+   "Arousal": {
+    "n": 300,
+    "r": 0.36988546538143463,
+    "rho": 0.3387522083578706,
+    "mae_fixed": 2.2696427422856527,
+    "rmse_fixed": 2.8638542626801544,
+    "mae_oracle": 1.7590022151022042,
+    "rmse_oracle": 2.2116054603497415,
+    "pred_mean": 1.480083871682485,
+    "gold_mean": 3.8847222222222224
+   },
+   "Astonishment_Surprise": {
+    "n": 300,
+    "r": 0.3677483215730027,
+    "rho": 0.35717019077989753,
+    "mae_fixed": 3.7590078764255086,
+    "rmse_fixed": 4.402010446394356,
+    "mae_oracle": 2.0422447547388725,
+    "rmse_oracle": 2.4964532945028926,
+    "pred_mean": 1.1740328698853653,
+    "gold_mean": 5.302777777777778
+   },
+   "Authenticity": {
+    "n": 300,
+    "r": 0.5773345684264289,
+    "rho": 0.46435004833387034,
+    "mae_fixed": 2.534679525239127,
+    "rmse_fixed": 3.033081308538925,
+    "mae_oracle": 2.1890828086358827,
+    "rmse_oracle": 2.5775022504868543,
+    "pred_mean": 2.8405809672673543,
+    "gold_mean": 5.2375
+   },
+   "Awe": {
+    "n": 300,
+    "r": 0.1768657368435447,
+    "rho": 0.15251236124845832,
+    "mae_fixed": 3.2818835085819638,
+    "rmse_fixed": 3.8772123396366505,
+    "mae_oracle": 1.9477386658852072,
+    "rmse_oracle": 2.4244842244008686,
+    "pred_mean": 0.887364858587583,
+    "gold_mean": 4.293055555555556
+   },
+   "Bitterness": {
+    "n": 300,
+    "r": 0.35965217351717443,
+    "rho": 0.3299663329592551,
+    "mae_fixed": 4.84372289838299,
+    "rmse_fixed": 5.408985178295415,
+    "mae_oracle": 1.996835629107499,
+    "rmse_oracle": 2.4250231635148025,
+    "pred_mean": 0.9597182302673658,
+    "gold_mean": 6.184722222222222
+   },
+   "Concentration": {
+    "n": 300,
+    "r": 0.27696334738090034,
+    "rho": 0.2761421793575484,
+    "mae_fixed": 6.0196387774414495,
+    "rmse_fixed": 6.5159051766525184,
+    "mae_oracle": 2.10340734277475,
+    "rmse_oracle": 2.4978301056127155,
+    "pred_mean": 0.6921464099486669,
+    "gold_mean": 7.011111111111113
+   },
+   "Confusion": {
+    "n": 300,
+    "r": 0.36718854645067056,
+    "rho": 0.3638884876498628,
+    "mae_fixed": 4.409959189853971,
+    "rmse_fixed": 5.050592214291727,
+    "mae_oracle": 2.0900024437237903,
+    "rmse_oracle": 2.572910854698658,
+    "pred_mean": 0.8864845130840937,
+    "gold_mean": 5.598611111111111
+   },
+   "Contemplation": {
+    "n": 300,
+    "r": 0.2620559214204224,
+    "rho": 0.23827331414793498,
+    "mae_fixed": 4.424059057424939,
+    "rmse_fixed": 4.991067863568061,
+    "mae_oracle": 1.9664163222230104,
+    "rmse_oracle": 2.3910640168578214,
+    "pred_mean": 0.9320303199688593,
+    "gold_mean": 5.7125
+   },
+   "Contempt": {
+    "n": 300,
+    "r": 0.46090538580596246,
+    "rho": 0.45337748197202193,
+    "mae_fixed": 4.437360869655533,
+    "rmse_fixed": 5.006909187369414,
+    "mae_oracle": 1.958644141444015,
+    "rmse_oracle": 2.394330749573155,
+    "pred_mean": 1.0450135364135107,
+    "gold_mean": 5.863888888888888
+   },
+   "Contentment": {
+    "n": 300,
+    "r": 0.2844950977241904,
+    "rho": 0.3030287003188924,
+    "mae_fixed": 4.236333942082194,
+    "rmse_fixed": 4.971633634789142,
+    "mae_oracle": 2.4156604848170042,
+    "rmse_oracle": 2.8463633795024132,
+    "pred_mean": 0.9948122915625572,
+    "gold_mean": 5.497222222222222
+   },
+   "Disappointment": {
+    "n": 300,
+    "r": 0.30942625558569486,
+    "rho": 0.28948721652462805,
+    "mae_fixed": 4.002372980732766,
+    "rmse_fixed": 4.636941175545593,
+    "mae_oracle": 2.3219560511497734,
+    "rmse_oracle": 2.771733710766294,
+    "pred_mean": 1.512597906490167,
+    "gold_mean": 5.875
+   },
+   "Disgust": {
+    "n": 300,
+    "r": 0.4574644991612823,
+    "rho": 0.45978421982466466,
+    "mae_fixed": 4.106396085165796,
+    "rmse_fixed": 4.798330966644815,
+    "mae_oracle": 2.080160648100575,
+    "rmse_oracle": 2.507845206165613,
+    "pred_mean": 0.9061448602875074,
+    "gold_mean": 5.294444444444445
+   },
+   "Distress": {
+    "n": 300,
+    "r": 0.4946328891365716,
+    "rho": 0.49704730052556134,
+    "mae_fixed": 3.9852916757738783,
+    "rmse_fixed": 4.648804641542442,
+    "mae_oracle": 2.1623412268103097,
+    "rmse_oracle": 2.538847063200711,
+    "pred_mean": 1.3712088053425153,
+    "gold_mean": 5.809722222222223
+   },
+   "Doubt": {
+    "n": 300,
+    "r": 0.31530384058830546,
+    "rho": 0.3241471571906354,
+    "mae_fixed": 4.187202382371538,
+    "rmse_fixed": 4.8508261279059495,
+    "mae_oracle": 2.137178597023092,
+    "rmse_oracle": 2.636815576404857,
+    "pred_mean": 0.7416112877925237,
+    "gold_mean": 5.1
+   },
+   "Elation": {
+    "n": 300,
+    "r": 0.438454605309588,
+    "rho": 0.44557872865254056,
+    "mae_fixed": 3.7355396309069224,
+    "rmse_fixed": 4.413566244285708,
+    "mae_oracle": 2.1669046445530755,
+    "rmse_oracle": 2.631097477627744,
+    "pred_mean": 0.9946924361586571,
+    "gold_mean": 4.954166666666667
+   },
+   "Embarrassment": {
+    "n": 300,
+    "r": 0.37959707434258805,
+    "rho": 0.3787464305158946,
+    "mae_fixed": 5.038938848650645,
+    "rmse_fixed": 5.774929206647175,
+    "mae_oracle": 2.4442419768685846,
+    "rmse_oracle": 2.9069029026814404,
+    "pred_mean": 0.6149268480141957,
+    "gold_mean": 5.830555555555556
+   },
+   "Emotional_Numbness": {
+    "n": 300,
+    "r": 0.1710212483507571,
+    "rho": 0.21743308258980656,
+    "mae_fixed": 2.6023798967637712,
+    "rmse_fixed": 3.1314473695976934,
+    "mae_oracle": 2.0746062907939455,
+    "rmse_oracle": 2.4252211208689105,
+    "pred_mean": 0.9201529668768247,
+    "gold_mean": 3.2902777777777774
+   },
+   "Fatigue_Exhaustion": {
+    "n": 300,
+    "r": 0.39704702663966945,
+    "rho": 0.4283465371837465,
+    "mae_fixed": 3.6452303323717343,
+    "rmse_fixed": 4.335837842202908,
+    "mae_oracle": 2.0816405587415385,
+    "rmse_oracle": 2.558878983072941,
+    "pred_mean": 0.7796556250751019,
+    "gold_mean": 4.5625
+   },
+   "Fear": {
+    "n": 300,
+    "r": 0.41296938120537624,
+    "rho": 0.4021698018866876,
+    "mae_fixed": 3.583116494781441,
+    "rmse_fixed": 4.347893879667345,
+    "mae_oracle": 2.29658888174792,
+    "rmse_oracle": 2.7726545730806094,
+    "pred_mean": 0.8748160953819751,
+    "gold_mean": 4.505555555555555
+   },
+   "Helplessness": {
+    "n": 300,
+    "r": 0.4536521494233645,
+    "rho": 0.43846842742697134,
+    "mae_fixed": 3.6023151036765837,
+    "rmse_fixed": 4.323538011282245,
+    "mae_oracle": 2.1439081291088047,
+    "rmse_oracle": 2.609475100740385,
+    "pred_mean": 1.2347555940349897,
+    "gold_mean": 5.134722222222222
+   },
+   "Hope_Enthusiasm_Optimism": {
+    "n": 300,
+    "r": 0.10839892751476043,
+    "rho": 0.13416193513261257,
+    "mae_fixed": 3.2982541276348956,
+    "rmse_fixed": 3.959065914718382,
+    "mae_oracle": 2.095661055646836,
+    "rmse_oracle": 2.59606692329946,
+    "pred_mean": 1.1543883587916692,
+    "gold_mean": 4.531944444444445
+   },
+   "Impatience_and_Irritability": {
+    "n": 300,
+    "r": 0.6513425407710236,
+    "rho": 0.6404088934321491,
+    "mae_fixed": 3.6985188394311876,
+    "rmse_fixed": 4.322855342288069,
+    "mae_oracle": 1.8882932015357308,
+    "rmse_oracle": 2.336821605214232,
+    "pred_mean": 1.4329011170069377,
+    "gold_mean": 5.518055555555555
+   },
+   "Infatuation": {
+    "n": 300,
+    "r": 0.35777096372921474,
+    "rho": 0.37657351748352763,
+    "mae_fixed": 4.596122657164695,
+    "rmse_fixed": 5.172390122685306,
+    "mae_oracle": 2.0617752178723974,
+    "rmse_oracle": 2.5351703413781896,
+    "pred_mean": 1.108161357541879,
+    "gold_mean": 6.0625
+   },
+   "Interest": {
+    "n": 300,
+    "r": 0.2891455746990581,
+    "rho": 0.26130068111867905,
+    "mae_fixed": 4.084851580195957,
+    "rmse_fixed": 4.618582133899773,
+    "mae_oracle": 1.8133373756793485,
+    "rmse_oracle": 2.276498658561329,
+    "pred_mean": 1.0311764494578044,
+    "gold_mean": 5.491666666666666
+   },
+   "Intoxication_Altered_States_of_Consciousness": {
+    "n": 300,
+    "r": 0.34203027922391066,
+    "rho": 0.35680618673540815,
+    "mae_fixed": 3.6390646080885607,
+    "rmse_fixed": 4.43557440524187,
+    "mae_oracle": 2.2775324248266027,
+    "rmse_oracle": 2.76706377143682,
+    "pred_mean": 0.816928861985604,
+    "gold_mean": 4.6
+   },
+   "Jealousy_&_Envy": {
+    "n": 300,
+    "r": 0.4063655525792801,
+    "rho": 0.3810060111779019,
+    "mae_fixed": 4.632406755075568,
+    "rmse_fixed": 5.345637766941067,
+    "mae_oracle": 2.3460828389316855,
+    "rmse_oracle": 2.79866309496171,
+    "pred_mean": 0.9295893756548563,
+    "gold_mean": 5.819444444444445
+   },
+   "Longing": {
+    "n": 300,
+    "r": 0.3514571019995419,
+    "rho": 0.3632369248547206,
+    "mae_fixed": 3.4211021887641104,
+    "rmse_fixed": 4.035488292336046,
+    "mae_oracle": 1.918830245731815,
+    "rmse_oracle": 2.4158144247093905,
+    "pred_mean": 1.0835671686629453,
+    "gold_mean": 4.769444444444444
+   },
+   "Malevolence_Malice": {
+    "n": 300,
+    "r": 0.5817974682642449,
+    "rho": 0.5592719919110212,
+    "mae_fixed": 3.9887468109291695,
+    "rmse_fixed": 4.63019567412644,
+    "mae_oracle": 2.015183091758302,
+    "rmse_oracle": 2.4865648391174444,
+    "pred_mean": 1.3995813056329887,
+    "gold_mean": 5.80138888888889
+   },
+   "Pain": {
+    "n": 300,
+    "r": 0.49895436450948843,
+    "rho": 0.4783924265825176,
+    "mae_fixed": 4.0566965674360596,
+    "rmse_fixed": 4.8087633546580735,
+    "mae_oracle": 2.2137541034204236,
+    "rmse_oracle": 2.7098815950151685,
+    "pred_mean": 0.8536707629263401,
+    "gold_mean": 5.043055555555556
+   },
+   "Pleasure_Ecstasy": {
+    "n": 300,
+    "r": 0.35488193793995443,
+    "rho": 0.3440611562350693,
+    "mae_fixed": 3.4912367623003697,
+    "rmse_fixed": 4.13925616692909,
+    "mae_oracle": 2.1198662724353214,
+    "rmse_oracle": 2.60402176737354,
+    "pred_mean": 0.9682265009482701,
+    "gold_mean": 4.594444444444445
+   },
+   "Pride": {
+    "n": 300,
+    "r": 0.3960479505887934,
+    "rho": 0.36805342281580905,
+    "mae_fixed": 4.088157255238011,
+    "rmse_fixed": 4.797526634766373,
+    "mae_oracle": 2.2537306691531422,
+    "rmse_oracle": 2.7247504538584297,
+    "pred_mean": 1.0652135128279527,
+    "gold_mean": 5.427777777777778
+   },
+   "Relief": {
+    "n": 300,
+    "r": 0.27679060813181655,
+    "rho": 0.26257758419537996,
+    "mae_fixed": 3.7019074294775254,
+    "rmse_fixed": 4.44130028550442,
+    "mae_oracle": 2.471305356978577,
+    "rmse_oracle": 2.9602372516661957,
+    "pred_mean": 1.4727657523751259,
+    "gold_mean": 5.406944444444445
+   },
+   "Sadness": {
+    "n": 300,
+    "r": 0.43992445753731674,
+    "rho": 0.38333892598806646,
+    "mae_fixed": 2.866232760345179,
+    "rmse_fixed": 3.6040618550210692,
+    "mae_oracle": 2.2295896877582395,
+    "rmse_oracle": 2.698238823603029,
+    "pred_mean": 1.5003521485626699,
+    "gold_mean": 4.4625
+   },
+   "Sexual_Lust": {
+    "n": 300,
+    "r": 0.4848176467410941,
+    "rho": 0.4565206280069779,
+    "mae_fixed": 3.272985048473827,
+    "rmse_fixed": 4.019107275377681,
+    "mae_oracle": 2.16176955491655,
+    "rmse_oracle": 2.619045395644244,
+    "pred_mean": 0.8718536521991094,
+    "gold_mean": 4.188888888888888
+   },
+   "Shame": {
+    "n": 300,
+    "r": 0.33547103654032573,
+    "rho": 0.3020629118101312,
+    "mae_fixed": 4.449012202769518,
+    "rmse_fixed": 5.198605916766589,
+    "mae_oracle": 2.415643556394109,
+    "rmse_oracle": 2.85517628318112,
+    "pred_mean": 0.8940592830131451,
+    "gold_mean": 5.591666666666667
+   },
+   "Sourness": {
+    "n": 300,
+    "r": 0.3403520763934405,
+    "rho": 0.3500745563839598,
+    "mae_fixed": 4.452306704696208,
+    "rmse_fixed": 4.992031036730612,
+    "mae_oracle": 1.9433154734318419,
+    "rmse_oracle": 2.359535333278849,
+    "pred_mean": 0.9978329380849997,
+    "gold_mean": 5.808333333333334
+   },
+   "Teasing": {
+    "n": 300,
+    "r": 0.6516931517518402,
+    "rho": 0.6431218124645829,
+    "mae_fixed": 3.713464834813088,
+    "rmse_fixed": 4.425790161733943,
+    "mae_oracle": 1.8895558955942184,
+    "rmse_oracle": 2.3374444596890043,
+    "pred_mean": 0.9931111877163251,
+    "gold_mean": 4.998611111111111
+   },
+   "Thankfulness_Gratitude": {
+    "n": 300,
+    "r": 0.30936060207043453,
+    "rho": 0.25008188979877555,
+    "mae_fixed": 3.084051876243145,
+    "rmse_fixed": 3.6768907349983206,
+    "mae_oracle": 2.178685934593257,
+    "rmse_oracle": 2.69595231417195,
+    "pred_mean": 1.92301530217131,
+    "gold_mean": 5.245833333333334
+   },
+   "Triumph": {
+    "n": 300,
+    "r": 0.27366532515134007,
+    "rho": 0.2898565539617107,
+    "mae_fixed": 4.42685413973199,
+    "rmse_fixed": 5.138883235669358,
+    "mae_oracle": 2.4126771578279067,
+    "rmse_oracle": 2.9037960893987496,
+    "pred_mean": 1.1585489691793918,
+    "gold_mean": 5.891666666666667
+   }
+  },
+  "macro": {
+   "r": 0.3863468032071092,
+   "rho": 0.3744752084546971,
+   "mae_fixed": 3.88378247795704,
+   "rmse_fixed": 4.534688176956223,
+   "mae_oracle": 2.1239726616032737,
+   "rmse_oracle": 2.5785113199231953,
+   "n_classes": 42,
+   "n_rows": 12600,
+   "pooled_r": 0.3035581904968801,
+   "pooled_rho": 0.32346719057601503
+  }
+ },
+ "unanimous": {
+  "per_class": {
+   "Affection": {
+    "n": 60,
+    "r": 0.5529150717110402,
+    "rho": 0.43378716310086146,
+    "mae_fixed": 5.51755503600552,
+    "rmse_fixed": 6.156692117402714,
+    "mae_oracle": 2.732385806363998,
+    "rmse_oracle": 3.426392446733921,
+    "pred_mean": 1.5056589022278786,
+    "gold_mean": 7.083333333333333
+   },
+   "Amusement": {
+    "n": 64,
+    "r": 0.7994487125873442,
+    "rho": 0.7261446886446885,
+    "mae_fixed": 3.980388536251017,
+    "rmse_fixed": 5.183139029653672,
+    "mae_oracle": 2.293531033019552,
+    "rmse_oracle": 2.7809162426078684,
+    "pred_mean": 1.0969131318852305,
+    "gold_mean": 5.234375
+   },
+   "Anger": {
+    "n": 92,
+    "r": 0.7200178403187404,
+    "rho": 0.6179893038238059,
+    "mae_fixed": 4.864459035698301,
+    "rmse_fixed": 5.504351290084606,
+    "mae_oracle": 2.4503440692663934,
+    "rmse_oracle": 3.0143681064425247,
+    "pred_mean": 1.851415508305249,
+    "gold_mean": 6.8478260869565215
+   },
+   "Arousal": {
+    "n": 48,
+    "r": 0.5524006485889821,
+    "rho": 0.6223404255319149,
+    "mae_fixed": 2.710940083932309,
+    "rmse_fixed": 3.625237112700168,
+    "mae_oracle": 2.7402702275726813,
+    "rmse_oracle": 3.393086392638652,
+    "pred_mean": 1.3409036882221699,
+    "gold_mean": 2.8125
+   },
+   "Astonishment_Surprise": {
+    "n": 54,
+    "r": 0.4876350412739648,
+    "rho": 0.4887745378311415,
+    "mae_fixed": 5.144944209506902,
+    "rmse_fixed": 6.016664392210057,
+    "mae_oracle": 3.2175548540932835,
+    "rmse_oracle": 3.866715011628534,
+    "pred_mean": 1.283237336134469,
+    "gold_mean": 6.296296296296297
+   },
+   "Authenticity": {
+    "n": 86,
+    "r": 0.8185642436988404,
+    "rho": 0.6024529458936743,
+    "mae_fixed": 4.265371058074342,
+    "rmse_fixed": 4.460687013074761,
+    "mae_oracle": 1.7074091344357831,
+    "rmse_oracle": 2.8689666350453638,
+    "pred_mean": 2.6258605080981585,
+    "gold_mean": 4.767441860465116
+   },
+   "Awe": {
+    "n": 50,
+    "r": 0.2812673910462358,
+    "rho": 0.1568787515006002,
+    "mae_fixed": 2.846779237048967,
+    "rmse_fixed": 4.015106725123012,
+    "mae_oracle": 3.1000313546194307,
+    "rmse_oracle": 3.7462438652798387,
+    "pred_mean": 0.8902920147776604,
+    "gold_mean": 2.6
+   },
+   "Bitterness": {
+    "n": 59,
+    "r": 0.513815347517619,
+    "rho": 0.28866160140268854,
+    "mae_fixed": 7.026446891033043,
+    "rmse_fixed": 7.521332322008985,
+    "mae_oracle": 2.258255199965875,
+    "rmse_oracle": 2.8045070293498626,
+    "pred_mean": 1.0019212622763747,
+    "gold_mean": 8.305084745762711
+   },
+   "Concentration": {
+    "n": 90,
+    "r": 0.33433365354689776,
+    "rho": 0.23663003663003662,
+    "mae_fixed": 7.872660077753521,
+    "rmse_fixed": 8.246597885433195,
+    "mae_oracle": 1.7026383210254505,
+    "rmse_oracle": 2.4536023741625512,
+    "pred_mean": 0.7887893153561486,
+    "gold_mean": 9.0
+   },
+   "Confusion": {
+    "n": 57,
+    "r": 0.6232633703851844,
+    "rho": 0.5045372050816698,
+    "mae_fixed": 5.419474142535885,
+    "rmse_fixed": 6.523002797547713,
+    "mae_oracle": 2.9907814286746754,
+    "rmse_oracle": 3.533938817477698,
+    "pred_mean": 0.8323267154526293,
+    "gold_mean": 6.228070175438597
+   },
+   "Contemplation": {
+    "n": 54,
+    "r": 0.5465069019360529,
+    "rho": 0.5624928530588907,
+    "mae_fixed": 5.245484329720654,
+    "rmse_fixed": 5.950937701351039,
+    "mae_oracle": 2.389184469297782,
+    "rmse_oracle": 2.9516965615831583,
+    "pred_mean": 0.9652680732585766,
+    "gold_mean": 6.388888888888889
+   },
+   "Contempt": {
+    "n": 60,
+    "r": 0.6664057189320528,
+    "rho": 0.5284245623784386,
+    "mae_fixed": 5.85079151391983,
+    "rmse_fixed": 6.609151626465388,
+    "mae_oracle": 2.550162664128132,
+    "rmse_oracle": 3.0716308926308615,
+    "pred_mean": 1.1472915545105935,
+    "gold_mean": 7.166666666666667
+   },
+   "Contentment": {
+    "n": 71,
+    "r": 0.45226434282404565,
+    "rho": 0.41703554661301134,
+    "mae_fixed": 4.833521236716142,
+    "rmse_fixed": 5.986295869525734,
+    "mae_oracle": 3.5533626909175493,
+    "rmse_oracle": 4.068532031996588,
+    "pred_mean": 0.9455380267660383,
+    "gold_mean": 5.563380281690141
+   },
+   "Disappointment": {
+    "n": 76,
+    "r": 0.46531036520164165,
+    "rho": 0.38411483253588513,
+    "mae_fixed": 5.28395692376714,
+    "rmse_fixed": 5.986026763781796,
+    "mae_oracle": 3.341776134513859,
+    "rmse_oracle": 3.9336647123999406,
+    "pred_mean": 1.5332392089461024,
+    "gold_mean": 6.5131578947368425
+   },
+   "Disgust": {
+    "n": 61,
+    "r": 0.7009981263954133,
+    "rho": 0.6598096245372816,
+    "mae_fixed": 5.583669682855629,
+    "rmse_fixed": 6.540534796539658,
+    "mae_oracle": 2.443780474199391,
+    "rmse_oracle": 3.1717426682845717,
+    "pred_mean": 0.9289430311468782,
+    "gold_mean": 6.39344262295082
+   },
+   "Distress": {
+    "n": 65,
+    "r": 0.6134174577985223,
+    "rho": 0.5968094405594405,
+    "mae_fixed": 5.487703765486622,
+    "rmse_fixed": 6.123337254428481,
+    "mae_oracle": 2.661292342685861,
+    "rmse_oracle": 3.2466109121973377,
+    "pred_mean": 1.5483444796158716,
+    "gold_mean": 7.153846153846154
+   },
+   "Doubt": {
+    "n": 63,
+    "r": 0.3815275914184077,
+    "rho": 0.3812884024577573,
+    "mae_fixed": 4.721351008431442,
+    "rmse_fixed": 5.934835283463176,
+    "mae_oracle": 3.747320545119541,
+    "rmse_oracle": 4.157738484939771,
+    "pred_mean": 0.7394334470468854,
+    "gold_mean": 5.079365079365079
+   },
+   "Elation": {
+    "n": 62,
+    "r": 0.6054671043012002,
+    "rho": 0.5149454811009544,
+    "mae_fixed": 4.047661371494768,
+    "rmse_fixed": 5.2901702571401445,
+    "mae_oracle": 3.1997280427109804,
+    "rmse_oracle": 3.8106599116803825,
+    "pred_mean": 1.0677154487179172,
+    "gold_mean": 4.758064516129032
+   },
+   "Embarrassment": {
+    "n": 86,
+    "r": 0.4070257670066185,
+    "rho": 0.40808528704184155,
+    "mae_fixed": 6.2755699089397226,
+    "rmse_fixed": 7.289926164690695,
+    "mae_oracle": 3.3540515370468262,
+    "rmse_oracle": 4.003755277384168,
+    "pred_mean": 0.6226667716752651,
+    "gold_mean": 6.8604651162790695
+   },
+   "Emotional_Numbness": {
+    "n": 80,
+    "r": 0.13927173161516246,
+    "rho": 0.05185185185185185,
+    "mae_fixed": 1.4888783879578114,
+    "rmse_fixed": 2.1175191746635202,
+    "mae_oracle": 0.8177006770620967,
+    "rmse_oracle": 1.9483345795892637,
+    "pred_mean": 0.8456336300820112,
+    "gold_mean": 0.4375
+   },
+   "Fatigue_Exhaustion": {
+    "n": 69,
+    "r": 0.5864975799156386,
+    "rho": 0.5723419802703691,
+    "mae_fixed": 3.6069729918031714,
+    "rmse_fixed": 4.856005872996595,
+    "mae_oracle": 2.7154203070805414,
+    "rmse_oracle": 3.4383753763873477,
+    "pred_mean": 0.7869844168856523,
+    "gold_mean": 3.9130434782608696
+   },
+   "Fear": {
+    "n": 85,
+    "r": 0.6749840846000648,
+    "rho": 0.5955442642173149,
+    "mae_fixed": 3.538074775653727,
+    "rmse_fixed": 4.936510038954487,
+    "mae_oracle": 2.7897998512654043,
+    "rmse_oracle": 3.4240770362396034,
+    "pred_mean": 0.8367363016395007,
+    "gold_mean": 3.588235294117647
+   },
+   "Helplessness": {
+    "n": 72,
+    "r": 0.6312506516629683,
+    "rho": 0.673290886873754,
+    "mae_fixed": 4.716602549370792,
+    "rmse_fixed": 5.682134146384557,
+    "mae_oracle": 3.1372205318948128,
+    "rmse_oracle": 3.6077922599308296,
+    "pred_mean": 1.205471518759926,
+    "gold_mean": 5.486111111111111
+   },
+   "Hope_Enthusiasm_Optimism": {
+    "n": 55,
+    "r": 0.15149885155367923,
+    "rho": 0.21327561327561334,
+    "mae_fixed": 3.3451195490824706,
+    "rmse_fixed": 4.359545340715086,
+    "mae_oracle": 3.674514685699086,
+    "rmse_oracle": 4.052665515340747,
+    "pred_mean": 1.1411360773173245,
+    "gold_mean": 3.090909090909091
+   },
+   "Impatience_and_Irritability": {
+    "n": 86,
+    "r": 0.8063730262607269,
+    "rho": 0.6894759186754091,
+    "mae_fixed": 5.036284531964811,
+    "rmse_fixed": 5.722364083270066,
+    "mae_oracle": 2.178312966298656,
+    "rmse_oracle": 2.6909848222500785,
+    "pred_mean": 1.6728902767563976,
+    "gold_mean": 6.686046511627907
+   },
+   "Infatuation": {
+    "n": 63,
+    "r": 0.36680538913094723,
+    "rho": 0.3531586021505376,
+    "mae_fixed": 6.474343046985247,
+    "rmse_fixed": 7.106763442303195,
+    "mae_oracle": 3.043535750092653,
+    "rmse_oracle": 3.786424271304742,
+    "pred_mean": 1.1318745258308591,
+    "gold_mean": 7.5396825396825395
+   },
+   "Interest": {
+    "n": 46,
+    "r": 0.4848783850505931,
+    "rho": 0.5696577243293247,
+    "mae_fixed": 4.685330472377516,
+    "rmse_fixed": 5.60384542670069,
+    "mae_oracle": 2.828156856578471,
+    "rmse_oracle": 3.407721277781528,
+    "pred_mean": 0.9822243037431136,
+    "gold_mean": 5.760869565217392
+   },
+   "Intoxication_Altered_States_of_Consciousness": {
+    "n": 68,
+    "r": 0.4462726308736932,
+    "rho": 0.44188265831965484,
+    "mae_fixed": 4.111093198489241,
+    "rmse_fixed": 5.495086269370943,
+    "mae_oracle": 3.6610221816753232,
+    "rmse_oracle": 4.1963456804530885,
+    "pred_mean": 0.7955027947969296,
+    "gold_mean": 4.338235294117647
+   },
+   "Jealousy_&_Envy": {
+    "n": 88,
+    "r": 0.5555433457022662,
+    "rho": 0.3731111972103836,
+    "mae_fixed": 6.210675046041415,
+    "rmse_fixed": 7.022735823261174,
+    "mae_oracle": 2.9523784113977856,
+    "rmse_oracle": 3.5926895718159595,
+    "pred_mean": 1.0357414573769679,
+    "gold_mean": 7.215909090909091
+   },
+   "Longing": {
+    "n": 52,
+    "r": 0.5440455851622302,
+    "rho": 0.47178348843165707,
+    "mae_fixed": 3.8181158363982863,
+    "rmse_fixed": 5.015544104432697,
+    "mae_oracle": 3.242504855169389,
+    "rmse_oracle": 3.834560089074945,
+    "pred_mean": 0.9993230709089682,
+    "gold_mean": 4.134615384615385
+   },
+   "Malevolence_Malice": {
+    "n": 81,
+    "r": 0.7873714941646293,
+    "rho": 0.6154245709123759,
+    "mae_fixed": 4.965503124858547,
+    "rmse_fixed": 5.785801473068402,
+    "mae_oracle": 2.096979139520237,
+    "rmse_oracle": 2.8339558193259973,
+    "pred_mean": 1.4857998376275285,
+    "gold_mean": 6.419753086419753
+   },
+   "Pain": {
+    "n": 86,
+    "r": 0.676280308982103,
+    "rho": 0.6552290202368035,
+    "mae_fixed": 4.563232578114418,
+    "rmse_fixed": 5.773716941654529,
+    "mae_oracle": 2.9227875851099725,
+    "rmse_oracle": 3.484907277413776,
+    "pred_mean": 0.9065182423868845,
+    "gold_mean": 5.058139534883721
+   },
+   "Pleasure_Ecstasy": {
+    "n": 60,
+    "r": 0.5516204523315417,
+    "rho": 0.48802445123645466,
+    "mae_fixed": 3.8431124829110646,
+    "rmse_fixed": 5.067391013876058,
+    "mae_oracle": 3.2918073939757244,
+    "rmse_oracle": 3.9923219519567703,
+    "pred_mean": 1.0577796002229054,
+    "gold_mean": 4.083333333333333
+   },
+   "Pride": {
+    "n": 70,
+    "r": 0.6250214852768,
+    "rho": 0.5868078033417898,
+    "mae_fixed": 5.189060988778971,
+    "rmse_fixed": 6.243483817948482,
+    "mae_oracle": 3.0956681334588345,
+    "rmse_oracle": 3.749056312718068,
+    "pred_mean": 1.0608959962214743,
+    "gold_mean": 5.928571428571429
+   },
+   "Relief": {
+    "n": 81,
+    "r": 0.45185063520729013,
+    "rho": 0.3966350496838302,
+    "mae_fixed": 5.155131819492085,
+    "rmse_fixed": 5.942410704898899,
+    "mae_oracle": 3.7320410703275875,
+    "rmse_oracle": 4.269774099553561,
+    "pred_mean": 1.4407750033302071,
+    "gold_mean": 5.925925925925926
+   },
+   "Sadness": {
+    "n": 79,
+    "r": 0.6504101184318947,
+    "rho": 0.5429892891918209,
+    "mae_fixed": 3.8144568837572823,
+    "rmse_fixed": 4.681623580564307,
+    "mae_oracle": 2.9168789818760397,
+    "rmse_oracle": 3.579562005076107,
+    "pred_mean": 1.491366845520237,
+    "gold_mean": 4.2405063291139244
+   },
+   "Sexual_Lust": {
+    "n": 83,
+    "r": 0.5959634207679974,
+    "rho": 0.5563368456404013,
+    "mae_fixed": 2.6532488456076058,
+    "rmse_fixed": 4.107428196750301,
+    "mae_oracle": 2.6364123088539992,
+    "rmse_oracle": 3.377890520938296,
+    "pred_mean": 0.7420724109712854,
+    "gold_mean": 2.5301204819277108
+   },
+   "Shame": {
+    "n": 72,
+    "r": 0.5066403217103982,
+    "rho": 0.3316933564859477,
+    "mae_fixed": 5.856606125358551,
+    "rmse_fixed": 6.831570525666321,
+    "mae_oracle": 3.319094923409996,
+    "rmse_oracle": 3.9237895946817027,
+    "pred_mean": 0.9109360939926572,
+    "gold_mean": 6.597222222222222
+   },
+   "Sourness": {
+    "n": 43,
+    "r": 0.5389620046706932,
+    "rho": 0.5767139836907279,
+    "mae_fixed": 5.831538818009272,
+    "rmse_fixed": 6.55965142751311,
+    "mae_oracle": 2.598758239427996,
+    "rmse_oracle": 3.401398458881026,
+    "pred_mean": 1.0561438403850378,
+    "gold_mean": 6.8604651162790695
+   },
+   "Teasing": {
+    "n": 84,
+    "r": 0.7615478628766329,
+    "rho": 0.682636428065202,
+    "mae_fixed": 4.1097957904444264,
+    "rmse_fixed": 5.296248413468334,
+    "mae_oracle": 2.3036460078950385,
+    "rmse_oracle": 2.9156316177011736,
+    "pred_mean": 0.9533230885863304,
+    "gold_mean": 5.0
+   },
+   "Thankfulness_Gratitude": {
+    "n": 64,
+    "r": 0.5046342260844684,
+    "rho": 0.46103479853479845,
+    "mae_fixed": 4.293885913211852,
+    "rmse_fixed": 4.900017986791032,
+    "mae_oracle": 3.5054373660717597,
+    "rmse_oracle": 4.073206035982343,
+    "pred_mean": 1.9044717724900693,
+    "gold_mean": 5.078125
+   },
+   "Triumph": {
+    "n": 87,
+    "r": 0.4407382647158411,
+    "rho": 0.4694721877961653,
+    "mae_fixed": 6.244111047638657,
+    "rmse_fixed": 6.934790733726855,
+    "mae_oracle": 3.044509678089605,
+    "rmse_oracle": 3.8397746087488227,
+    "pred_mean": 1.19819836010193,
+    "gold_mean": 7.35632183908046
+   }
+  },
+  "macro": {
+   "r": 0.5476439655532633,
+   "rho": 0.4880851109558755,
+   "mae_fixed": 4.7745214965114045,
+   "rmse_fixed": 5.6906241652763,
+   "mae_oracle": 2.8318678150449537,
+   "rmse_oracle": 3.469666837085937,
+   "n_classes": 42,
+   "n_rows": 2912,
+   "pooled_r": 0.4450673682975008,
+   "pooled_rho": 0.43795912771501133
+  }
+ },
+ "presence_auc": {
+  "per_class": {
+   "Affection": {
+    "auc": 0.7893772893772893,
+    "n_absent": 13,
+    "n_present": 126
+   },
+   "Amusement": {
+    "auc": 0.9028548770816812,
+    "n_absent": 26,
+    "n_present": 97
+   },
+   "Anger": {
+    "auc": 0.8755707762557078,
+    "n_absent": 24,
+    "n_present": 146
+   },
+   "Arousal": {
+    "auc": 0.8232258064516129,
+    "n_absent": 31,
+    "n_present": 50
+   },
+   "Astonishment_Surprise": {
+    "auc": 0.74375,
+    "n_absent": 16,
+    "n_present": 110
+   },
+   "Authenticity": {
+    "auc": 0.9386752136752137,
+    "n_absent": 45,
+    "n_present": 104
+   },
+   "Awe": {
+    "auc": 0.6627384960718294,
+    "n_absent": 33,
+    "n_present": 54
+   },
+   "Confusion": {
+    "auc": 0.8691148775894538,
+    "n_absent": 18,
+    "n_present": 118
+   },
+   "Contempt": {
+    "auc": 0.8951048951048951,
+    "n_absent": 13,
+    "n_present": 121
+   },
+   "Contentment": {
+    "auc": 0.7863559797866367,
+    "n_absent": 26,
+    "n_present": 137
+   },
+   "Disappointment": {
+    "auc": 0.7630094043887148,
+    "n_absent": 22,
+    "n_present": 145
+   },
+   "Disgust": {
+    "auc": 0.865009746588694,
+    "n_absent": 18,
+    "n_present": 114
+   },
+   "Distress": {
+    "auc": 0.8291505791505791,
+    "n_absent": 14,
+    "n_present": 148
+   },
+   "Doubt": {
+    "auc": 0.7345454545454545,
+    "n_absent": 25,
+    "n_present": 110
+   },
+   "Elation": {
+    "auc": 0.8199346405228758,
+    "n_absent": 30,
+    "n_present": 102
+   },
+   "Embarrassment": {
+    "auc": 0.7524154589371981,
+    "n_absent": 23,
+    "n_present": 144
+   },
+   "Emotional_Numbness": {
+    "auc": 0.7543859649122807,
+    "n_absent": 76,
+    "n_present": 15
+   },
+   "Fatigue_Exhaustion": {
+    "auc": 0.8356777493606138,
+    "n_absent": 34,
+    "n_present": 92
+   },
+   "Fear": {
+    "auc": 0.8358798735511064,
+    "n_absent": 52,
+    "n_present": 73
+   },
+   "Helplessness": {
+    "auc": 0.8564560439560439,
+    "n_absent": 28,
+    "n_present": 104
+   },
+   "Hope_Enthusiasm_Optimism": {
+    "auc": 0.5687645687645687,
+    "n_absent": 33,
+    "n_present": 78
+   },
+   "Impatience_and_Irritability": {
+    "auc": 0.9395604395604396,
+    "n_absent": 26,
+    "n_present": 112
+   },
+   "Infatuation": {
+    "auc": 0.7819628647214855,
+    "n_absent": 13,
+    "n_present": 145
+   },
+   "Interest": {
+    "auc": 0.8429027113237639,
+    "n_absent": 11,
+    "n_present": 114
+   },
+   "Intoxication_Altered_States_of_Consciousness": {
+    "auc": 0.7950437317784257,
+    "n_absent": 35,
+    "n_present": 98
+   },
+   "Jealousy_&_Envy": {
+    "auc": 0.7911931818181818,
+    "n_absent": 22,
+    "n_present": 128
+   },
+   "Longing": {
+    "auc": 0.8195631528964862,
+    "n_absent": 27,
+    "n_present": 78
+   },
+   "Malevolence_Malice": {
+    "auc": 0.9222527472527473,
+    "n_absent": 26,
+    "n_present": 140
+   },
+   "Pain": {
+    "auc": 0.8255360623781677,
+    "n_absent": 38,
+    "n_present": 108
+   },
+   "Pleasure_Ecstasy": {
+    "auc": 0.7454611474219317,
+    "n_absent": 34,
+    "n_present": 81
+   },
+   "Pride": {
+    "auc": 0.8106641123882503,
+    "n_absent": 27,
+    "n_present": 116
+   },
+   "Relief": {
+    "auc": 0.734832790766499,
+    "n_absent": 31,
+    "n_present": 109
+   },
+   "Sadness": {
+    "auc": 0.8195639701663798,
+    "n_absent": 42,
+    "n_present": 83
+   },
+   "Sexual_Lust": {
+    "auc": 0.8469907407407408,
+    "n_absent": 60,
+    "n_present": 72
+   },
+   "Shame": {
+    "auc": 0.7918112421929215,
+    "n_absent": 22,
+    "n_present": 131
+   },
+   "Teasing": {
+    "auc": 0.9382596013612057,
+    "n_absent": 34,
+    "n_present": 121
+   },
+   "Thankfulness_Gratitude": {
+    "auc": 0.7208387942332897,
+    "n_absent": 28,
+    "n_present": 109
+   },
+   "Triumph": {
+    "auc": 0.7600299962504687,
+    "n_absent": 21,
+    "n_present": 127
+   }
+  },
+  "macro_auc": 0.8102227627190484,
+  "n_classes": 38
+ }
+}

bench/emonet/emonet_pred_target.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eae32166c0f5b49524576b70f0a86df665829b010beeb71e1441971e2c8c816d
+size 100928

bench/emonet/eval_emonet.py

@@ -0,0 +1,164 @@
+#!/usr/bin/env python3
+"""Evaluate the 61 VoiceClap attribute-regression heads on EmoNet Voice Bench.
+
+Protocol (matches arXiv:2506.09827 Table 5): each row has ONE target emotion and
+a human intensity in {0,1,2} averaged over 2-4 experts, mapped to 0/5/10.
+We take the single head named by the row's emotion and score the clip with it.
+Primary metrics are the scale-free ones (Pearson r, Spearman rho) because our
+heads emit the Empathic-Insight 0-7 scale, not 0-10.
+"""
+import io, json, os, sys, time
+import numpy as np
+import pyarrow.parquet as pq
+import soundfile as sf
+import torch
+import torch.nn.functional as F
+
+BENCH = os.path.dirname(os.path.abspath(__file__))
+NB = "/e/data1/datasets/playground/mmlaion/schuhmann1/dramabox"
+sys.path.insert(0, f"{NB}/headspub/repo")
+from voiceclap_heads import AttributeScorer  # noqa: E402
+
+SR, MAXS = 16000, 480000
+OUT = f"{BENCH}/results"
+os.makedirs(OUT, exist_ok=True)
+
+
+def decode(b):
+    with sf.SoundFile(io.BytesIO(b)) as f:
+        sr = f.samplerate
+        x = f.read(frames=int(30.0 * sr), dtype="float32", always_2d=True)
+    x = x.mean(1) if x.shape[1] > 1 else x[:, 0]
+    if len(x) < 400:
+        return None
+    if sr != SR:
+        import torchaudio
+        x = torchaudio.functional.resample(torch.from_numpy(np.ascontiguousarray(x)), sr, SR).numpy()
+    return x[:MAXS]
+
+
+def pearson(a, b):
+    if a.std() < 1e-9 or b.std() < 1e-9:
+        return float("nan")
+    return float(np.corrcoef(a, b)[0, 1])
+
+
+def spearman(a, b):
+    return pearson(np.argsort(np.argsort(a)).astype(float), np.argsort(np.argsort(b)).astype(float))
+
+
+def main():
+    lab = pq.read_table(f"{BENCH}/labels.parquet").to_pydict()
+    n = len(lab["clip_id"])
+    print(f"{n} benchmark rows", flush=True)
+
+    scorer = AttributeScorer(heads_path=f"{NB}/attrdistill/heads/heads.pt")
+    print("device", scorer.device, flush=True)
+
+    preds = np.full(n, np.nan, dtype=np.float64)
+    all61 = np.full((n, 61), np.nan, dtype=np.float32)
+    dims61 = scorer.dims
+
+    # group row indices by source parquet so each file is opened once
+    by_file = {}
+    for i in range(n):
+        by_file.setdefault(lab["parquet_file"][i], []).append(i)
+
+    t0 = time.time()
+    done = 0
+    for fpath, idxs in by_file.items():
+        tbl = pq.read_table(os.path.join(BENCH, fpath), columns=["audioId"])
+        col = tbl.column("audioId")
+        CH = 128
+        for s in range(0, len(idxs), CH):
+            chunk = idxs[s:s + CH]
+            wavs, keep = [], []
+            for i in chunk:
+                rec = col[lab["row_index"][i]].as_py()
+                w = decode(rec["bytes"])
+                if w is None:
+                    continue
+                wavs.append(w)
+                keep.append(i)
+            if not wavs:
+                continue
+            emb = scorer.embed(wavs, batch_size=32)
+            sc = scorer.score_embeddings(emb)
+            M = np.stack([sc[d] for d in dims61], 1)
+            for j, i in enumerate(keep):
+                all61[i] = M[j]
+                preds[i] = sc[lab["head_name"][i]][j]
+            done += len(chunk)
+            if done % 1280 < CH:
+                print(f"  {done}/{n}  {time.time()-t0:.0f}s", flush=True)
+
+    gold = np.asarray(lab["gold_0_10"], dtype=np.float64)
+    head = np.asarray(lab["head_name"])
+    ok = np.isfinite(preds)
+    print(f"scored {ok.sum()}/{n}", flush=True)
+
+    def block(mask, name):
+        per = {}
+        for c in sorted(set(head[mask])):
+            m = mask & (head == c)
+            if m.sum() < 20:
+                continue
+            p, g = preds[m], gold[m]
+            # oracle affine (upper bound on MAE/RMSE, fit ON the benchmark - reported as such)
+            A = np.stack([p, np.ones_like(p)], 1)
+            coef, *_ = np.linalg.lstsq(A, g, rcond=None)
+            pc = A @ coef
+            per[c] = {
+                "n": int(m.sum()), "r": pearson(p, g), "rho": spearman(p, g),
+                "mae_fixed": float(np.abs(np.clip(p / 7.0 * 10.0, 0, 10) - g).mean()),
+                "rmse_fixed": float(np.sqrt(((np.clip(p / 7.0 * 10.0, 0, 10) - g) ** 2).mean())),
+                "mae_oracle": float(np.abs(pc - g).mean()),
+                "rmse_oracle": float(np.sqrt(((pc - g) ** 2).mean())),
+                "pred_mean": float(p.mean()), "gold_mean": float(g.mean()),
+            }
+        agg = {k: float(np.nanmean([v[k] for v in per.values()]))
+               for k in ["r", "rho", "mae_fixed", "rmse_fixed", "mae_oracle", "rmse_oracle"]}
+        agg["n_classes"] = len(per)
+        agg["n_rows"] = int(mask.sum())
+        pooled_p, pooled_g = preds[mask], gold[mask]
+        agg["pooled_r"] = pearson(pooled_p, pooled_g)
+        agg["pooled_rho"] = spearman(pooled_p, pooled_g)
+        print(f"\n=== {name}: rows={agg['n_rows']} classes={agg['n_classes']} "
+              f"macro r={agg['r']:.3f} rho={agg['rho']:.3f} "
+              f"MAE(fixed)={agg['mae_fixed']:.2f} MAE(oracle)={agg['mae_oracle']:.2f}", flush=True)
+        return {"per_class": per, "macro": agg}
+
+    res = {"all": block(ok, "ALL 12,600 rows")}
+    una = np.asarray(lab["unanimous"], dtype=bool)
+    res["unanimous"] = block(ok & una, "unanimous-annotator subset")
+
+    # presence detection: absent-by-all (gold==0) vs present-by-all
+    pa = np.asarray(lab["present_all"], dtype=bool)
+    absent = ok & (gold == 0.0)
+    present = ok & pa
+    aucs = {}
+    for c in sorted(set(head[ok])):
+        a = preds[absent & (head == c)]
+        b = preds[present & (head == c)]
+        if len(a) < 10 or len(b) < 10:
+            continue
+        allv = np.concatenate([a, b])
+        r = np.argsort(np.argsort(allv)).astype(float) + 1
+        n1 = len(b)
+        auc = (r[len(a):].sum() - n1 * (n1 + 1) / 2) / (len(a) * n1)
+        aucs[c] = {"auc": float(auc), "n_absent": len(a), "n_present": n1}
+    res["presence_auc"] = {"per_class": aucs,
+                           "macro_auc": float(np.mean([v["auc"] for v in aucs.values()])) if aucs else None,
+                           "n_classes": len(aucs)}
+    print(f"\npresence-detection macro ROC-AUC over {len(aucs)} classes: "
+          f"{res['presence_auc']['macro_auc']:.3f}", flush=True)
+
+    json.dump(res, open(f"{OUT}/emonet_metrics.json", "w"), indent=1)
+    np.save(f"{OUT}/emonet_pred_target.npy", preds)
+    np.save(f"{OUT}/emonet_pred_all61.f16.npy", all61.astype(np.float16))
+    json.dump(dims61, open(f"{OUT}/emonet_dims61.json", "w"))
+    print("WROTE", OUT, flush=True)
+
+
+if __name__ == "__main__":
+    main()