Add files using upload-large-folder tool

.gitattributes

@@ -1,60 +1,3 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.avro filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.lz4 filter=lfs diff=lfs merge=lfs -text
-*.mds filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
-# Audio files - uncompressed
-*.pcm filter=lfs diff=lfs merge=lfs -text
-*.sam filter=lfs diff=lfs merge=lfs -text
-*.raw filter=lfs diff=lfs merge=lfs -text
-# Audio files - compressed
-*.aac filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
 *.flac filter=lfs diff=lfs merge=lfs -text
 *.mp3 filter=lfs diff=lfs merge=lfs -text
-*.ogg filter=lfs diff=lfs merge=lfs -text
-*.wav filter=lfs diff=lfs merge=lfs -text
-# Image files - uncompressed
-*.bmp filter=lfs diff=lfs merge=lfs -text
-*.gif filter=lfs diff=lfs merge=lfs -text
-*.png filter=lfs diff=lfs merge=lfs -text
-*.tiff filter=lfs diff=lfs merge=lfs -text
-# Image files - compressed
-*.jpg filter=lfs diff=lfs merge=lfs -text
-*.jpeg filter=lfs diff=lfs merge=lfs -text
-*.webp filter=lfs diff=lfs merge=lfs -text
-# Video files - compressed
-*.mp4 filter=lfs diff=lfs merge=lfs -text
-*.webm filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,115 @@
+---
+license: cc-by-nc-4.0
+language:
+- en
+pretty_name: VoxParadox
+task_categories:
+- audio-classification
+- question-answering
+tags:
+- audio
+- speech
+- paralinguistic
+- benchmark
+- adversarial
+- audio-llm
+size_categories:
+- 1K<n<10K
+configs:
+- config_name: default
+  data_files:
+  - split: test
+    path: metadata.jsonl
+---
+
+# VoxParadox
+
+An adversarial speech QA benchmark for **paralinguistic understanding** in
+Audio LLMs. Each example is built around a controlled linguistic–acoustic
+contradiction: the transcript explicitly asserts an incorrect paralinguistic
+attribute, while the audio reliably conveys the correct one. Models that
+defer to transcript content are misled; models that listen are not.
+
+**2,000 MCQs** across **10 paralinguistic tasks** (200 each).
+
+## Quick start
+
+```python
+from datasets import load_dataset
+ds = load_dataset("IHP-Lab/VoxParadox", split="test")
+print(ds[0])           # includes `audio` (decoded), `question`, `choice_a..d`,
+                       # `answer_gt`, `adversarial_labels`, `task_name`, `id`
+```
+
+## Tasks
+
+`y_true` (audio) and `y_adv` (transcript) are disjoint by construction.
+
+| Task (`task_name`) | Acoustic attribute the model must recover |
+|---|---|
+| `age_prediction` | Speaker's age group |
+| `gender_prediction` | Speaker's gender |
+| `emotion_recognition` | Emotion conveyed by delivery (high-contrast pairs) |
+| `intonation_perception` | Rising vs. falling intonation |
+| `speaker_identity_recognition` | Which segment shares a speaker with a queried segment |
+| `total_speaker_counting` | Number of distinct speakers |
+| `pitch_comparison` | Ordering of three segments by pitch |
+| `volume_comparison` | Ordering of three segments by loudness |
+| `speed_comparison` | Ordering of three segments by speaking rate |
+| `vocal_range_comparison` | Ordering of three segments by pitch range |
+
+## File layout
+
+```
+VoxParadox_public_release/
+├── metadata.jsonl    # one record per example (loaded by `datasets`)
+├── voxparadox.json   # same content as JSON array (for direct inspection)
+├── audio/            # 2,000 wav files
+└── eval.py           # evaluation script (GT accuracy + ALA)
+```
+
+## Record schema
+
+| Field | Type | Description |
+|---|---|---|
+| `id` | string | `{task_name}__{N}`, with `N` running 0–199 within each task. |
+| `task_name` | string | One of the 10 tasks above. |
+| `file_name` / `audio_path` | string | Path to the audio clip, relative to this directory. |
+| `question` | string | The MCQ question prompt. |
+| `choice_a` / `choice_b` / `choice_c` / `choice_d` | string | The four answer options. |
+| `answer_gt` | string | The acoustic ground-truth `y_true` (one of the four choices). |
+| `adversarial_labels` | list[string] | The transcript-implied label(s) `y_adv`. Single-element for most tasks; 2 elements for the four `*_comparison` tasks. |
+
+## Evaluation
+
+Two complementary metrics:
+
+* **GT Accuracy** — fraction matching `answer_gt`. Higher is better; reflects use of acoustic evidence.
+* **Adversarial-Label Agreement (ALA)** — fraction matching any string in `adversarial_labels`. Higher ALA means more transcript-following.
+
+Run `eval.py` on a JSONL of model predictions (one record per line, fields `id` and `response`):
+
+```bash
+python eval.py --predictions preds.jsonl
+```
+
+Example prediction record:
+```json
+{"id": "age_prediction__0", "response": "(C) Elderly adult."}
+```
+
+The script parses A/B/C/D from the response (letter-first, then choice-text
+fallback), prints per-task and overall GT/ALA, and optionally writes a JSON
+report with `--report report.json`.
+
+## License
+
+Released under **CC BY-NC 4.0**. Audio was synthesized via commercial TTS
+engines (ElevenLabs, GPT-4o, Microsoft Azure); commercial reuse of the audio
+is subject to those vendors' terms of service.
+
+## Citation
+
+> *Do Audio LLMs Listen or Read? Analyzing and Mitigating Paralinguistic Failures with VoxParadox.*
+
+(BibTeX to be added upon publication.)

audio/age_prediction__150.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4dfd5a72a9c4bbcdd6dde6da5147592632a179aa2631e8858f66355195334d73
+size 102584

audio/age_prediction__151.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70ff8b988872e1b1c77c2615a6b3f9be607144d16f552a150767a4181f0778b0
+size 160540

audio/age_prediction__171.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3dcc30ba4eb5358b38ad387f0c1778f1b9e1a5502de126b459a262f624fe515
+size 83264

audio/age_prediction__172.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0f33bf45677a19975782ed1182853f710f019591e5a0f2f9bb409e2458499b0
+size 104070

audio/age_prediction__199.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34c97941c569a4907496668435e84dfe144d9b12b355c9bcc5c20996c06ee985
+size 144194

audio/emotion_recognition__102.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f16de844008ddb481e7b28d59dab2199ed184c35addae86e89ba46be5e6f9e07
+size 171342

audio/emotion_recognition__132.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00602d3468447c58af022c16d3fd4360610d0141e9e9b70ec17a60cc087b38b7
+size 126798

audio/emotion_recognition__170.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5fb586d71fed9972fb96ab73c84f1f231b99cd5ba71c39e945dbc853bbb6d4a
+size 149070

audio/emotion_recognition__40.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5249d22aef0b8011314c5379ac4003f77f187b4c444fac2027e4104406a421ba
+size 215118

audio/emotion_recognition__85.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebd1dc5f8eea7ac04f4031e5f4a35250af8860521889862d429c06f9c795b333
+size 134478

audio/emotion_recognition__92.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:861fc732378a15710f846886b824e7092eb6abb37d6e1af43fe4f84e563d47c1
+size 178254

audio/gender_prediction__103.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4c69249c0590dc14134a510364102eb408aa45af2ab0f528c4c43d01a6a754b
+size 108528

audio/gender_prediction__104.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b32603091839c8334dcb58ff1d8f1d414ab1bc86dfe7580e7f14eee1d71d9b0
+size 92182

audio/gender_prediction__151.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65f2b0820dc7d51d91a36614354b64a9ff68c9cdc6057fe920fb93bcb9f9cc00
+size 110014

audio/gender_prediction__184.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97b31e961171741f157bf76cc6ceea5c526da402a8520a461812d3d16f606608
+size 86236

audio/gender_prediction__31.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b92cf7b79dc94e2a91c2e0eb6161a1148e3483437253ee20933ce68a1b347bbb
+size 111500

audio/intonation_perception__117.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:985a73dcbae5b6f8462ad467aec1024f819b651465f387736bcdc9fc6255442c
+size 115646

audio/intonation_perception__157.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3df54b67deb34bef394dd340e23663b59df606048614a977951447025bb4408
+size 103646

audio/intonation_perception__34.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81d94a1f6b64b731848cbe1ee239a471f6b2233c41bdca6b578915ae8f47d21e
+size 138046

audio/intonation_perception__36.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:854dfefc9c1881b5918e6c8cee31d5edc017e7f95eab52e9a1c0a8b97de97373
+size 102846

audio/intonation_perception__45.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6636bdee7313265e714584c7f916c013a43b2f05b84d1ef959f1f8a33baf3ce0
+size 103246

audio/intonation_perception__77.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0489e1a5bd91780f53f5a583a62116afa6f46c7131bfb1b57f8634e7cf3e9c04
+size 88446

audio/intonation_perception__97.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4ab9d2795cccc4beebea2f150cbe2ee496b88e6329359ba1af1ad3c164ae213
+size 120046

audio/pitch_comparison__11.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78308fe48c9e7b3e7cdd182c86c745c901c7c5d641f0d919a20d50f3aca6bd86
+size 598220

audio/pitch_comparison__165.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0d0a2becea8435fb25eb0797fd3288fed7b105c4c355f84ec55118c5ddb9c76
+size 370028

audio/pitch_comparison__172.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3ed0b2740ec4dc1ecf63e0fe583c52202729578823abf0e7a28796fa736782b
+size 449804

audio/pitch_comparison__24.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cde27cf78c59ae446723ec70a6ae3be7edb6eea3a5f930f1643b271d1d7c00af
+size 490988

audio/pitch_comparison__52.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db488e97d2b9271d4e6037a3d7b4842f05f2c6d3d5b665ec0f48c577d284a58f
+size 511820

audio/speaker_identity_recognition__145.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57e0218b3ba6d23d807afa0a7968b20f62334289a88b18176d68bfbd7374a761
+size 899060

audio/speaker_identity_recognition__32.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad8df9724373d2680ec46154ed3304d8e1e6390f8e490af0029d46753a5cfacb
+size 787316

audio/speaker_identity_recognition__7.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa22c4c172883efcbcdbcf82a0511c638b91047d021b3cd974f9d966bd7ee9f9
+size 920948

audio/speaker_identity_recognition__85.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bedc399f897d52abd6ce4767231170589f7806e870a1b4accf4f13d2969b086f
+size 941684

audio/speed_comparison__129.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa44a703f330bb96517ae52e32cc63b8c31bafc6f706d184182988de38e387c2
+size 279830

audio/speed_comparison__189.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39b4fa31e030dd0a19e65c9c3268f1ad634c080b2e027c2bebfd774a9c480c71
+size 352054

audio/total_speaker_counting__101.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e35f49f15b938f1fb3e04202b59cf2af96e13b905c57676757fbdd99b536036
+size 421180

audio/total_speaker_counting__134.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:089dd7ae7b7ce1bc404fea08995b4aba7ee1b6b91cf2cd60cb01ffe55750312c
+size 414268

audio/total_speaker_counting__157.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57069dbf3dbd3333509fb5e90f7de74b791dec19eb19233f9c5ebb75dcddfea7
+size 584764

audio/total_speaker_counting__196.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2df538d2f0f9b25d60f362fd3e2fc1b09c341bc59c2b4a236ff8da11fbee12cd
+size 379708

audio/vocal_range_comparison__38.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e713ecfd020d2a301c89450e93d75b5d5bf3b3a36af3ba3d7dd853630455f89
+size 663788

audio/vocal_range_comparison__59.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00714896a5310edab88188595f96db7d0cae12cf2e8b0f80f3ba8ea94ba57e81
+size 518924

audio/volume_comparison__119.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25f35ebad88a8af6ad9d32879ed43cc7b03c4323b9cecc0ac0468eadbc8450d3
+size 397484

audio/volume_comparison__156.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:01a1b3afe1dd18e88110c4767b7ce506819061ebfcbdf92e5ee2dfcb11c38dd2
+size 338732

audio/volume_comparison__29.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98ff1c95ec674c5c5a414c6ebbb3c736731ce5165e9221e0aaa66b3216f7d33f
+size 432044

audio/volume_comparison__66.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c7c3f20aff560697c89771ab1cb41b8a512cf4792f7bc197b4f9fba3a0377d0
+size 763820

audio/volume_comparison__79.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23a89fbf7cc01535e167f5d094bcc86a0808cd41036fdc048b29de366274f3bb
+size 259244

eval.py

@@ -0,0 +1,244 @@
+"""Evaluate model predictions on VoxParadox.
+
+Usage:
+    python eval.py --predictions <preds.jsonl>
+    python eval.py --predictions <preds.jsonl> --dataset voxparadox.json --report report.json
+
+Predictions file format (JSONL, one JSON object per line):
+    {"id": "age_prediction__0", "response": "Elderly adult"}
+
+The `response` field is the raw model output string. The script parses it into
+one of the four MCQ choices (A/B/C/D) using letter-extraction heuristics with a
+choice-text fallback, then scores it against:
+
+  * GT Accuracy -- match rate against `answer_gt`.
+  * Adversarial-Label Agreement (ALA) -- match rate against any string in
+    `adversarial_labels` (the transcript-implied labels).
+
+Both metrics are reported per task and overall (macro = micro since each task
+has exactly 200 examples).
+"""
+
+import argparse
+import json
+import os
+import re
+import unicodedata
+from collections import defaultdict
+
+# ---------- Response parsing ----------
+
+MAX_TAIL_LINES = 2
+MAX_TAIL_CHARS = 1000
+
+
+def _nfkc(s):
+    return unicodedata.normalize("NFKC", s or "")
+
+
+def _norm(s):
+    s = _nfkc(str(s)).casefold().replace("–", "-").replace("—", "-")
+    return re.sub(r"\s+", " ", s).strip()
+
+
+def _tail(text):
+    text = _nfkc(text).replace("\r\n", "\n").replace("\r", "\n")
+    lines = [ln.strip() for ln in text.split("\n") if ln.strip()]
+    out = "\n".join(lines[-MAX_TAIL_LINES:]) if lines else text.strip()
+    return out[-MAX_TAIL_CHARS:] if len(out) > MAX_TAIL_CHARS else out
+
+
+def get_choices(item):
+    return {L: str(item.get(f"choice_{L.lower()}", "")) for L in "ABCD"}
+
+
+def _parse_choice_letter(response):
+    """Extract an A/B/C/D letter from the response tail.
+
+    First tries to match the entire last line as a single letter (possibly
+    with brackets/punctuation), e.g. "A", "(B)", "C." -- this is the most
+    confident signal. If that fails, scans the tail for any isolated A/B/C/D
+    mention not embedded inside a longer word, and returns the LAST such
+    occurrence (handles outputs like "I think the answer is B.").
+    """
+    if not response:
+        return None
+    tail = _tail(response)
+    lines = [ln.strip() for ln in tail.split("\n") if ln.strip()]
+    last = lines[-1] if lines else tail.strip()
+    m = re.match(r"(?i)^[\(\[\{]?\s*([ABCD])\s*[\)\]\}]?\s*[,.;:!?\-–—]*\s*$", last)
+    if m:
+        return m.group(1).upper()
+    rx = re.compile(r"(?i)(?<![A-Za-z0-9])[\(\[\{]?\s*([ABCD])\s*[\)\]\}]?\s*[,.;:!?\-–—]*\s*(?=$|\s)")
+    ms = list(rx.finditer(tail))
+    if ms:
+        return ms[-1].group(1).upper()
+    return None
+
+
+def _parse_choice_by_content(response, choices):
+    """Fallback parser: match by choice text appearing in the response tail.
+
+    Picks the choice whose normalized text appears LATEST in the tail,
+    ranked by `rfind` position. This mirrors the matching used to produce
+    the paper's reported numbers, including the known caveat that overlapping
+    choice text (e.g., "male" inside "female") is decided by position alone.
+    """
+    if not response:
+        return None
+    tail_n = _norm(_tail(response))
+    tail_n_sp = tail_n.replace("-", " ")
+    best_L, best_pos = None, -1
+    for L, txt in choices.items():
+        txt_n = _norm(txt)
+        if not txt_n:
+            continue
+        txt_n_sp = txt_n.replace("-", " ")
+        for t, hay in [(txt_n, tail_n), (txt_n_sp, tail_n_sp)]:
+            pos = hay.rfind(t)
+            if pos > best_pos:
+                best_pos = pos
+                best_L = L
+    return best_L if best_pos >= 0 else None
+
+
+def parse_response(response, item):
+    """Map a model response to one of A/B/C/D.
+
+    Letter-first: if the response contains an isolated A/B/C/D mention, use
+    it. Otherwise fall back to matching the choice text in the response tail.
+    This matches the parsing semantics used to produce the paper's reported
+    numbers, so results from this script are directly comparable.
+    """
+    L = _parse_choice_letter(response)
+    if L is not None:
+        return L
+    return _parse_choice_by_content(response, get_choices(item))
+
+
+# ---------- Label resolution ----------
+
+def text_to_letter(text, choices):
+    if not text:
+        return None
+    if text.upper() in {"A", "B", "C", "D"}:
+        return text.upper()
+    n = _norm(text)
+    for L, t in choices.items():
+        if _norm(t) == n:
+            return L
+    return None
+
+
+def gt_letter(item):
+    return text_to_letter(item.get("answer_gt", ""), get_choices(item))
+
+
+def adv_letters(item):
+    choices = get_choices(item)
+    out = set()
+    for s in item.get("adversarial_labels", []) or []:
+        L = text_to_letter(s, choices)
+        if L:
+            out.add(L)
+    return out
+
+
+# ---------- Evaluation ----------
+
+def evaluate(dataset, predictions):
+    pred_map = {p["id"]: p.get("response", "") for p in predictions}
+    gt_correct = defaultdict(int)
+    adv_correct = defaultdict(int)
+    total = defaultdict(int)
+    missing = 0
+    parse_fail = 0
+    for item in dataset:
+        task = item["task_name"]
+        total[task] += 1
+        resp = pred_map.get(item["id"])
+        if resp is None:
+            missing += 1
+            continue
+        pred = parse_response(resp, item)
+        if pred is None:
+            parse_fail += 1
+            continue
+        gt = gt_letter(item)
+        adv = adv_letters(item)
+        if gt and pred == gt:
+            gt_correct[task] += 1
+        if pred in adv:
+            adv_correct[task] += 1
+    return dict(gt_correct), dict(adv_correct), dict(total), missing, parse_fail
+
+
+def main():
+    ap = argparse.ArgumentParser(description="Evaluate model predictions on VoxParadox.")
+    ap.add_argument("--predictions", required=True, help="Path to predictions JSONL file.")
+    ap.add_argument("--dataset", default=os.path.join(os.path.dirname(__file__), "voxparadox.json"),
+                    help="Path to voxparadox.json (default: alongside this script).")
+    ap.add_argument("--report", default=None, help="Optional path to write a JSON report.")
+    args = ap.parse_args()
+
+    with open(args.dataset) as f:
+        dataset = json.load(f)
+    predictions = []
+    with open(args.predictions) as f:
+        for ln in f:
+            ln = ln.strip()
+            if not ln:
+                continue
+            predictions.append(json.loads(ln))
+
+    gt_c, adv_c, total, missing, parse_fail = evaluate(dataset, predictions)
+
+    tasks = sorted(total.keys())
+    n_total = sum(total.values())
+    sum_gt = sum(gt_c.get(t, 0) for t in tasks)
+    sum_adv = sum(adv_c.get(t, 0) for t in tasks)
+
+    print(f"VoxParadox Evaluation")
+    print(f"  Dataset:     {n_total} examples across {len(tasks)} tasks")
+    print(f"  Predictions: {len(predictions)} loaded "
+          f"(missing: {missing}, parse-failed: {parse_fail})")
+    print()
+    print(f"{'Task':<32} {'N':>5} {'GT Acc':>9} {'ALA':>9}")
+    print("-" * 58)
+    for t in tasks:
+        n = total[t]
+        gt = 100.0 * gt_c.get(t, 0) / n if n else 0
+        ala = 100.0 * adv_c.get(t, 0) / n if n else 0
+        print(f"{t:<32} {n:>5} {gt:>8.2f}% {ala:>8.2f}%")
+    print("-" * 58)
+    overall_gt = 100.0 * sum_gt / n_total if n_total else 0
+    overall_ala = 100.0 * sum_adv / n_total if n_total else 0
+    print(f"{'Overall':<32} {n_total:>5} {overall_gt:>8.2f}% {overall_ala:>8.2f}%")
+
+    if args.report:
+        report = {
+            "dataset": os.path.abspath(args.dataset),
+            "predictions": os.path.abspath(args.predictions),
+            "n_examples": n_total,
+            "n_predictions": len(predictions),
+            "n_missing": missing,
+            "n_parse_failed": parse_fail,
+            "overall": {"gt_acc": overall_gt, "ala": overall_ala},
+            "per_task": {
+                t: {
+                    "n": total[t],
+                    "gt_correct": gt_c.get(t, 0),
+                    "adv_correct": adv_c.get(t, 0),
+                    "gt_acc": 100.0 * gt_c.get(t, 0) / total[t] if total[t] else 0,
+                    "ala": 100.0 * adv_c.get(t, 0) / total[t] if total[t] else 0,
+                }
+                for t in tasks
+            },
+        }
+        with open(args.report, "w") as f:
+            json.dump(report, f, indent=2)
+        print(f"\n[report] {args.report}")
+
+
+if __name__ == "__main__":
+    main()