File size: 21,515 Bytes
be9aa9f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b3ee71e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
be9aa9f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b3ee71e
 
 
 
 
 
 
 
be9aa9f
b3ee71e
 
 
be9aa9f
 
b3ee71e
 
 
 
 
 
 
 
 
 
 
 
 
 
be9aa9f
b3ee71e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
be9aa9f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b3ee71e
 
be9aa9f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b3ee71e
be9aa9f
 
b3ee71e
 
 
 
 
 
be9aa9f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b3ee71e
be9aa9f
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
#!/usr/bin/env python3
"""
Voxtral ASR Fine-tuning Interface

Features:
- Collect a personal voice dataset (upload WAV/FLAC + transcripts or record mic audio)
- Build a JSONL dataset ({audio_path, text}) at 16kHz
- Fine-tune Voxtral (LoRA or full) with streamed logs
- Push model to Hugging Face Hub
- Deploy a Voxtral ASR demo Space

Env tokens (optional):
- HF_WRITE_TOKEN or HF_TOKEN: write access token
- HF_READ_TOKEN: optional read token
- HF_USERNAME: fallback username if not derivable from token
"""

from __future__ import annotations

import os
import json
from pathlib import Path
from datetime import datetime
from typing import Any, Dict, Generator, Optional, Tuple

import gradio as gr

PROJECT_ROOT = Path(__file__).resolve().parent


def get_python() -> str:
    import sys
    return sys.executable or "python"


def get_username_from_token(token: str) -> Optional[str]:
    try:
        from huggingface_hub import HfApi  # type: ignore
        api = HfApi(token=token)
        info = api.whoami()
        if isinstance(info, dict):
            return info.get("name") or info.get("username")
        if isinstance(info, str):
            return info
    except Exception:
        return None
    return None


def run_command_stream(args: list[str], env: Dict[str, str], cwd: Optional[Path] = None) -> Generator[str, None, int]:
    import subprocess
    import shlex
    yield f"$ {' '.join(shlex.quote(a) for a in ([get_python()] + args))}"
    process = subprocess.Popen(
        [get_python()] + args,
        stdout=subprocess.PIPE,
        stderr=subprocess.STDOUT,
        text=True,
        env=env,
        cwd=str(cwd or PROJECT_ROOT),
        bufsize=1,
        universal_newlines=True,
    )
    assert process.stdout is not None
    for line in iter(process.stdout.readline, ""):
        yield line.rstrip()
    process.stdout.close()
    code = process.wait()
    yield f"[exit_code={code}]"
    return code


def detect_nvidia_driver() -> Tuple[bool, str]:
    """Detect NVIDIA driver/GPU presence with multiple strategies.

    Returns (available, human_message).
    """
    # 1) Try torch CUDA
    try:
        import torch  # type: ignore
        if torch.cuda.is_available():
            try:
                num = torch.cuda.device_count()
                names = [torch.cuda.get_device_name(i) for i in range(num)]
                return True, f"NVIDIA GPU detected: {', '.join(names)}"
            except Exception:
                return True, "NVIDIA GPU detected (torch.cuda available)"
    except Exception:
        pass

    # 2) Try NVML via pynvml
    try:
        import pynvml  # type: ignore
        try:
            pynvml.nvmlInit()
            cnt = pynvml.nvmlDeviceGetCount()
            names = []
            for i in range(cnt):
                h = pynvml.nvmlDeviceGetHandleByIndex(i)
                names.append(pynvml.nvmlDeviceGetName(h).decode("utf-8", errors="ignore"))
            drv = pynvml.nvmlSystemGetDriverVersion().decode("utf-8", errors="ignore")
            pynvml.nvmlShutdown()
            if cnt > 0:
                return True, f"NVIDIA driver {drv}; GPUs: {', '.join(names)}"
        except Exception:
            pass
    except Exception:
        pass

    # 3) Try nvidia-smi
    try:
        import subprocess
        res = subprocess.run(["nvidia-smi", "-L"], capture_output=True, text=True, timeout=3)
        if res.returncode == 0 and res.stdout.strip():
            return True, res.stdout.strip().splitlines()[0]
    except Exception:
        pass

    return False, "No NVIDIA driver/GPU detected"


def duplicate_space_hint() -> str:
    space_id = os.environ.get("SPACE_ID") or os.environ.get("HF_SPACE_ID")
    if space_id:
        space_url = f"https://huggingface.co/spaces/{space_id}"
        dup_url = f"{space_url}?duplicate=true"
        return (
            f"ℹ️ No NVIDIA driver detected. If you're on Hugging Face Spaces, "
            f"please duplicate this Space to GPU hardware: [Duplicate this Space]({dup_url})."
        )
    return (
        "ℹ️ No NVIDIA driver detected. To enable training, run on a machine with an NVIDIA GPU/driver "
        "or duplicate this Space on Hugging Face with GPU hardware."
    )


def _write_jsonl(rows: list[dict], path: Path) -> Path:
    path.parent.mkdir(parents=True, exist_ok=True)
    with open(path, "w", encoding="utf-8") as f:
        for r in rows:
            f.write(json.dumps(r, ensure_ascii=False) + "\n")
    return path


def _save_uploaded_dataset(files: list, transcripts: list[str]) -> str:
    dataset_dir = PROJECT_ROOT / "datasets" / "voxtral_user"
    dataset_dir.mkdir(parents=True, exist_ok=True)
    rows: list[dict] = []
    for i, fpath in enumerate(files or []):
        if i >= len(transcripts):
            break
        rows.append({"audio_path": fpath, "text": transcripts[i] or ""})
    jsonl_path = dataset_dir / "data.jsonl"
    _write_jsonl(rows, jsonl_path)
    return str(jsonl_path)


def _save_recordings(recordings: list[tuple[int, list]], transcripts: list[str]) -> str:
    import soundfile as sf
    dataset_dir = PROJECT_ROOT / "datasets" / "voxtral_user"
    wav_dir = dataset_dir / "wavs"
    wav_dir.mkdir(parents=True, exist_ok=True)
    rows: list[dict] = []
    for i, rec in enumerate(recordings or []):
        if rec is None:
            continue
        if i >= len(transcripts):
            break
        sr, data = rec
        out_path = wav_dir / f"rec_{i:04d}.wav"
        sf.write(str(out_path), data, sr)
        rows.append({"audio_path": str(out_path), "text": transcripts[i] or ""})
    jsonl_path = dataset_dir / "data.jsonl"
    _write_jsonl(rows, jsonl_path)
    return str(jsonl_path)


def start_voxtral_training(
    use_lora: bool,
    base_model: str,
    repo_short: str,
    jsonl_path: str,
    train_count: int,
    eval_count: int,
    batch_size: int,
    grad_accum: int,
    learning_rate: float,
    epochs: float,
    lora_r: int,
    lora_alpha: int,
    lora_dropout: float,
    freeze_audio_tower: bool,
    push_to_hub: bool,
    deploy_demo: bool,
) -> Generator[str, None, None]:
    env = os.environ.copy()
    write_token = env.get("HF_WRITE_TOKEN") or env.get("HF_TOKEN")
    read_token = env.get("HF_READ_TOKEN")
    username = get_username_from_token(write_token or "") or env.get("HF_USERNAME") or ""
    output_dir = PROJECT_ROOT / "outputs" / repo_short

    # 1) Train
    script = PROJECT_ROOT / ("scripts/train_lora.py" if use_lora else "scripts/train.py")
    args = [str(script)]
    if jsonl_path:
        args += ["--dataset-jsonl", jsonl_path]
    args += [
        "--model-checkpoint", base_model,
        "--train-count", str(train_count),
        "--eval-count", str(eval_count),
        "--batch-size", str(batch_size),
        "--grad-accum", str(grad_accum),
        "--learning-rate", str(learning_rate),
        "--epochs", str(epochs),
        "--output-dir", str(output_dir),
        "--save-steps", "50",
    ]
    if use_lora:
        args += [
            "--lora-r", str(lora_r),
            "--lora-alpha", str(lora_alpha),
            "--lora-dropout", str(lora_dropout),
        ]
        if freeze_audio_tower:
            args += ["--freeze-audio-tower"]
    for line in run_command_stream(args, env):
        yield line

    # 2) Push to Hub
    if push_to_hub:
        repo_name = f"{username}/{repo_short}" if username else repo_short
        push_args = [
            str(PROJECT_ROOT / "scripts/push_to_huggingface.py"),
            str(output_dir),
            repo_name,
        ]
        for line in run_command_stream(push_args, env):
            yield line

    # 3) Deploy demo Space
    if deploy_demo and username:
        deploy_args = [
            str(PROJECT_ROOT / "scripts/deploy_demo_space.py"),
            "--hf-token", write_token or "",
            "--hf-username", username,
            "--model-id", f"{username}/{repo_short}",
            "--demo-type", "voxtral",
            "--space-name", f"{repo_short}-demo",
        ]
        for line in run_command_stream(deploy_args, env):
            yield line


def load_voxpopuli_phrases(language="en", max_phrases=None, split="train"):
    """Load phrases from VoxPopuli dataset.

    Args:
        language: Language code (e.g., 'en', 'de', 'fr', etc.)
        max_phrases: Maximum number of phrases to load (None for all available)
        split: Dataset split to use ('train', 'validation', 'test')

    Returns:
        List of normalized text phrases
    """
    try:
        from datasets import load_dataset
        import random

        # Load the specified language dataset
        ds = load_dataset("facebook/voxpopuli", language, split=split)

        # Extract normalized text phrases
        phrases = []
        for example in ds:
            text = example.get("normalized_text", "").strip()
            if text and len(text) > 10:  # Filter out very short phrases
                phrases.append(text)

        # Shuffle and limit if specified
        if max_phrases:
            phrases = random.sample(phrases, min(max_phrases, len(phrases)))
        else:
            # If no limit, shuffle the entire list
            random.shuffle(phrases)

        return phrases

    except Exception as e:
        print(f"Error loading VoxPopuli phrases: {e}")
        # Fallback to some basic phrases if loading fails
        return [
            "The quick brown fox jumps over the lazy dog.",
            "Please say your full name.",
            "Today is a good day to learn something new.",
            "Artificial intelligence helps with many tasks.",
            "I enjoy reading books and listening to music.",
        ]

# Initialize phrases dynamically
VOXPOPULI_LANGUAGE = "en"  # Default to English
ALL_PHRASES = load_voxpopuli_phrases(VOXPOPULI_LANGUAGE, max_phrases=None)

with gr.Blocks(title="Voxtral ASR Fine-tuning") as demo:
    has_gpu, gpu_msg = detect_nvidia_driver()
    if has_gpu:
        gr.HTML(
            f"""
            <div style="background-color: rgba(59, 130, 246, 0.1); border: 1px solid rgba(59, 130, 246, 0.3); border-radius: 8px; padding: 12px; margin-bottom: 16px; text-align: center;">
                <p style="color: rgb(59, 130, 246); margin: 0; font-size: 14px; font-weight: 600;">
                    ✅ NVIDIA GPU ready — {gpu_msg}
                </p>
                <p style="color: rgb(59, 130, 246); margin: 6px 0 0; font-size: 12px;">
                    Set HF_WRITE_TOKEN/HF_TOKEN in environment to enable Hub push.
                </p>
            </div>
            """
        )
    else:
        hint_md = duplicate_space_hint()
        gr.HTML(
            f"""
            <div style="background-color: rgba(245, 158, 11, 0.1); border: 1px solid rgba(245, 158, 11, 0.3); border-radius: 8px; padding: 12px; margin-bottom: 16px; text-align: center;">
                <p style="color: rgb(234, 88, 12); margin: 0; font-size: 14px; font-weight: 600;">
                    ⚠️ No NVIDIA GPU/driver detected — training requires a GPU runtime
                </p>
                <p style="color: rgb(234, 88, 12); margin: 6px 0 0; font-size: 12px;">
                    {hint_md}
                </p>
            </div>
            """
        )

    gr.Markdown("""
    # 🎙️ Voxtral ASR Fine-tuning
    Read the phrases below and record them. Then start fine-tuning.
    """)

    jsonl_out = gr.Textbox(label="Dataset JSONL path", interactive=False, visible=True)

    # Language selection for VoxPopuli phrases
    voxpopuli_lang = gr.Dropdown(
        choices=["en", "de", "fr", "es", "pl", "it", "ro", "hu", "cs", "nl", "fi", "hr", "sk", "sl", "et", "lt"],
        value="en",
        label="VoxPopuli Language",
        info="Select language for phrases from VoxPopuli dataset"
    )

    # Recording grid with dynamic text readouts
    phrase_texts_state = gr.State(ALL_PHRASES)
    visible_rows_state = gr.State(10)  # Start with 10 visible rows
    max_rows = len(ALL_PHRASES)  # No cap on total rows
    phrase_markdowns: list[gr.Markdown] = []
    rec_components = []

    def create_recording_grid(phrases, visible_count=10):
        """Create recording grid components dynamically"""
        markdowns = []
        recordings = []
        for idx, phrase in enumerate(phrases):
            visible = idx < visible_count
            md = gr.Markdown(f"**{idx+1}. {phrase}**", visible=visible)
            markdowns.append(md)
            comp = gr.Audio(sources="microphone", type="numpy", label=f"Recording {idx+1}", visible=visible)
            recordings.append(comp)
        return markdowns, recordings

    # Initial grid creation
    with gr.Column():
        phrase_markdowns, rec_components = create_recording_grid(ALL_PHRASES, 10)

    # Add more rows button
    add_rows_btn = gr.Button("➕ Add 10 More Rows", variant="secondary")

    def add_more_rows(current_visible, current_phrases):
        """Add 10 more rows by making them visible"""
        new_visible = min(current_visible + 10, len(current_phrases))
        visibility_updates = []
        for i in range(len(current_phrases)):
            if i < new_visible:
                visibility_updates.append(gr.update(visible=True))
            else:
                visibility_updates.append(gr.update(visible=False))
        return [new_visible] + visibility_updates

    def change_language(language):
        """Change the language and reload phrases from VoxPopuli"""
        new_phrases = load_voxpopuli_phrases(language, max_phrases=None)
        # Reset visible rows to 10
        visible_count = min(10, len(new_phrases))

        # Create combined updates for existing components (up to current length)
        current_len = len(phrase_markdowns)
        combined_updates = []

        # Update existing components
        for i in range(current_len):
            if i < len(new_phrases):
                if i < visible_count:
                    combined_updates.append(gr.update(value=f"**{i+1}. {new_phrases[i]}**", visible=True))
                else:
                    combined_updates.append(gr.update(visible=False))
            else:
                combined_updates.append(gr.update(visible=False))

        # If we have more phrases than components, we can't update them via Gradio
        # The interface will need to be reloaded for significantly different phrase counts
        return [new_phrases, visible_count] + combined_updates

    # Connect language change to phrase reloading
    voxpopuli_lang.change(
        change_language,
        inputs=[voxpopuli_lang],
        outputs=[phrase_texts_state, visible_rows_state] + phrase_markdowns + rec_components
    )

    add_rows_btn.click(
        add_more_rows,
        inputs=[visible_rows_state, phrase_texts_state],
        outputs=[visible_rows_state] + phrase_markdowns + rec_components
    )

    # Advanced options accordion
    with gr.Accordion("Advanced options", open=False):
        base_model = gr.Textbox(value="mistralai/Voxtral-Mini-3B-2507", label="Base Voxtral model")
        use_lora = gr.Checkbox(value=True, label="Use LoRA (parameter-efficient)")
        with gr.Row():
            batch_size = gr.Number(value=2, precision=0, label="Batch size")
            grad_accum = gr.Number(value=4, precision=0, label="Grad accum")
        with gr.Row():
            learning_rate = gr.Number(value=5e-5, precision=6, label="Learning rate")
            epochs = gr.Number(value=3.0, precision=2, label="Epochs")
        with gr.Accordion("LoRA settings", open=False):
            lora_r = gr.Number(value=8, precision=0, label="LoRA r")
            lora_alpha = gr.Number(value=32, precision=0, label="LoRA alpha")
            lora_dropout = gr.Number(value=0.0, precision=3, label="LoRA dropout")
            freeze_audio_tower = gr.Checkbox(value=True, label="Freeze audio tower")
        with gr.Row():
            train_count = gr.Number(value=100, precision=0, label="Train samples")
            eval_count = gr.Number(value=50, precision=0, label="Eval samples")
        repo_short = gr.Textbox(value=f"voxtral-finetune-{datetime.now().strftime('%Y%m%d_%H%M%S')}", label="Model repo (short)")
        push_to_hub = gr.Checkbox(value=True, label="Push to HF Hub after training")
        deploy_demo = gr.Checkbox(value=True, label="Deploy demo Space after push")

        gr.Markdown("### Upload audio + transcripts (optional)")
        upload_audio = gr.File(file_count="multiple", type="filepath", label="Upload WAV/FLAC files (optional)")
        transcripts_box = gr.Textbox(lines=6, label="Transcripts (one per line, aligned with files)")
        save_upload_btn = gr.Button("Save uploaded dataset")

        def _collect_upload(files, txt):
            lines = [s.strip() for s in (txt or "").splitlines() if s.strip()]
            return _save_uploaded_dataset(files or [], lines)

        save_upload_btn.click(_collect_upload, [upload_audio, transcripts_box], [jsonl_out])

    # Save recordings button
    save_rec_btn = gr.Button("Save recordings as dataset")

    def _collect_preloaded_recs(*recs_and_texts):
        import soundfile as sf
        dataset_dir = PROJECT_ROOT / "datasets" / "voxtral_user"
        wav_dir = dataset_dir / "wavs"
        wav_dir.mkdir(parents=True, exist_ok=True)
        rows: list[dict] = []
        if not recs_and_texts:
            jsonl_path = dataset_dir / "data.jsonl"
            _write_jsonl(rows, jsonl_path)
            return str(jsonl_path)
        texts = recs_and_texts[-1]
        recs = recs_and_texts[:-1]
        for i, rec in enumerate(recs):
            if rec is None:
                continue
            sr, data = rec
            out_path = wav_dir / f"rec_{i:04d}.wav"
            sf.write(str(out_path), data, sr)
            # Use the full phrase list (ALL_PHRASES) instead of just PHRASES
            label_text = (texts[i] if isinstance(texts, list) and i < len(texts) else (ALL_PHRASES[i] if i < len(ALL_PHRASES) else ""))
            rows.append({"audio_path": str(out_path), "text": label_text})
        jsonl_path = dataset_dir / "data.jsonl"
        _write_jsonl(rows, jsonl_path)
        return str(jsonl_path)

    save_rec_btn.click(_collect_preloaded_recs, rec_components + [phrase_texts_state], [jsonl_out])

    # Quick sample from VoxPopuli (few random rows)
    with gr.Row():
        vp_lang = gr.Dropdown(choices=["en", "de", "fr", "es", "it", "pl", "ro", "hu", "cs", "nl", "fi", "hr", "sk", "sl", "et", "lt"], value="en", label="VoxPopuli language")
        vp_samples = gr.Number(value=20, precision=0, label="Num samples")
        vp_split = gr.Dropdown(choices=["train", "validation", "test"], value="train", label="Split")
        vp_btn = gr.Button("Use VoxPopuli sample")

        def _collect_voxpopuli(lang_code: str, num_samples: int, split: str):
            import sys
            # Workaround for dill on Python 3.13 expecting __main__ during import
            if "__main__" not in sys.modules:
                sys.modules["__main__"] = sys.modules[__name__]
            from datasets import load_dataset, Audio  # type: ignore
            import random
            ds = load_dataset("facebook/voxpopuli", lang_code, split=split)
            ds = ds.cast_column("audio", Audio(sampling_rate=16000))
            # shuffle and select
            total = len(ds)
            k = max(1, min(int(num_samples or 1), total))
            ds = ds.shuffle(seed=random.randint(1, 10_000))
            ds_sel = ds.select(range(k))

            dataset_dir = PROJECT_ROOT / "datasets" / "voxtral_user"
            rows: list[dict] = []
            texts: list[str] = []
            for ex in ds_sel:
                audio = ex.get("audio") or {}
                path = audio.get("path")
                text = ex.get("normalized_text") or ex.get("raw_text") or ""
                if path and text is not None:
                    rows.append({"audio_path": path, "text": text})
                    texts.append(str(text))
            jsonl_path = dataset_dir / "data.jsonl"
            _write_jsonl(rows, jsonl_path)
            # Build markdown content updates for on-screen prompts
            combined_updates = []
            for i in range(len(phrase_markdowns)):
                t = texts[i] if i < len(texts) else ""
                if i < len(texts):
                    combined_updates.append(gr.update(value=f"**{i+1}. {t}**", visible=True))
                else:
                    combined_updates.append(gr.update(visible=False))

            return (str(jsonl_path), texts, *combined_updates)

        vp_btn.click(
            _collect_voxpopuli,
            [vp_lang, vp_samples, vp_split],
            [jsonl_out, phrase_texts_state] + phrase_markdowns,
        )

    start_btn = gr.Button("Start Fine-tuning")
    logs_box = gr.Textbox(label="Logs", lines=20)

    start_btn.click(
        start_voxtral_training,
        inputs=[
            use_lora, base_model, repo_short, jsonl_out, train_count, eval_count,
            batch_size, grad_accum, learning_rate, epochs,
            lora_r, lora_alpha, lora_dropout, freeze_audio_tower,
            push_to_hub, deploy_demo,
        ],
        outputs=[logs_box],
    )


if __name__ == "__main__":
    server_port = int(os.environ.get("INTERFACE_PORT", "7860"))
    server_name = os.environ.get("INTERFACE_HOST", "0.0.0.0")
    demo.queue().launch(server_name=server_name, server_port=server_port, mcp_server=True, ssr_mode=False)