captioning-whisper-large-turbo-wip

⚠️ Work in progress. This is a mid-training checkpoint uploaded while the full 3-epoch run is still going. A final checkpoint will replace these weights once training completes.

Fine-tune of openai/whisper-large-v3-turbo repurposed for audio captioning (not just speech transcription). The model takes 30s of audio and emits a long, descriptive natural-language caption covering speech, music, sound events, vocal bursts, and non-speech audio.

Checkpoint

• Step: 561,118 (out of an estimated 822,054 total steps over 3 epochs)
• Samples seen: ~35.9M (≈ 2.05 epochs)
• Best validation loss so far: ~0.679 (held-out set of 100 pairs per dataset, 896 total after divisibility trim)
• Precision: bf16
• Global batch size: 64 (8 GPUs × per-device 8)

Training setup

Hardware

• 8× A100/H100-class GPUs (80 GB each), single node, DDP

Hyperparameters

• learning_rate: 5e-4, cosine schedule, 5% warmup
• weight_decay: 0.0
• max_grad_norm: 1.0
• per_device_train_batch_size: 8
• gradient_accumulation_steps: 1
• bf16: true (fp16 was discarded because of grad-scaler overflows)
• max_audio_seconds: 30
• max_label_tokens: 448

Datasets (round-robin, each batch draws from all 9)

• laion/majestrino-data — speech + detailed captions
• laion/captioned-ai-music-snippets — music with comprehensive captions
• mitermix/audioset-with-grounded-captions — general audio events
• TTS-AGI/majestrino-unified-detailed-captions-temporal — speech with temporal/emotion descriptions
• laion/laions_got_talent_clean_with_captions — performance speech/music
• laion/freesound-commercially-permissive-subset-with-captions
• laion/generated-sound-events
• laion/in-the-wild-sound-events
• laion/synthetic_vocal_bursts

Total ~2.83M train pairs, 100 held out per dataset for validation.

Inference

import torch, librosa
from transformers import WhisperProcessor, WhisperForConditionalGeneration

model_id = "laion/captioning-whisper-large-turbo-wip"
processor = WhisperProcessor.from_pretrained(model_id)
model = WhisperForConditionalGeneration.from_pretrained(
    model_id, torch_dtype=torch.bfloat16
).to("cuda").eval()

audio, sr = librosa.load("clip.wav", sr=16000, mono=True, duration=30.0)
feats = processor.feature_extractor(
    audio, sampling_rate=16000, return_tensors="pt"
).input_features.to("cuda", torch.bfloat16)

with torch.no_grad():
    out = model.generate(
        feats,
        max_new_tokens=448,
        num_beams=4,
        do_sample=False,
    )
caption = processor.batch_decode(out, skip_special_tokens=True)[0]
print(caption)

The model emits long, multi-sentence descriptions. Keep max_new_tokens near 448 and prefer beam search over greedy for quality.

What's in this repo

• Model weights (model.safetensors) — bf16 Whisper-large-v3-turbo checkpoint at step 561,118
• Processor / tokenizer files — copied from the base model (vocab unchanged)
• code/ — the exact training pipeline used:
  • train.py — Seq2SeqTrainer-based fine-tuner with custom round-robin streaming dataset, manual rank-0 eval (HF distributed eval deadlocks here), and live monitor integration
  • prefetcher.py — tar-level background downloader that keeps ~2 tars per dataset hot on disk, with age-aware LRU eviction (MIN_EVICT_AGE_S = 600) to prevent in-flight eviction races
  • monitor.py — HTTP dashboard (port 8077) with live loss, eval, audio playback
  • watchdog.sh — auto-restart wrapper
  • download_data.py — one-shot HF → local extractor with post-extract delete
  • phase2_launcher.sh — phase-2 launch helper

Caveats

• No optimizer / scheduler state is included. The training script saves via save_pretrained only. Resuming optimizer-state-aware training from this checkpoint is not possible without starting a fresh optimizer (what we did for the in-flight resume: reduced peak LR to 1e-4 with a short 1% warmup since the cosine schedule had already decayed most of the way).
• No RNG / data-iterator state. The streaming dataset is round-robin across tars and workers but resuming won't replay the exact same sample order.
• This is mid-training. Expect captions to still be underfit on rarer datasets and to sometimes leak artifacts from base-model speech transcription behaviors.
• Languages: English-heavy. The base model is multilingual, but the fine-tuning data is mostly English.

Training code commit / state

The code/ directory in this repo is a snapshot of the exact scripts running at the time of this checkpoint. The Seq2SeqTrainer in transformers>=4.57 requires a few API tweaks (processing_class= not tokenizer=, eval_strategy= not evaluation_strategy=, _get_train_sampler(train_dataset) etc.) — those are wired up in train.py.

License

Apache 2.0 (inherited from base model; verify downstream dataset licenses before redistribution).