Bolo models v1

Trained models for Bolo (Apache-2.0, like the code). Download with uv run python scripts/get_models.py (default stack) or --all.

File	What it is
bolo-whisper-v1-fp16.zip	Bolo Sun v1, the default speech model (Swift fine-tuned on synthetic Hinglish command speech with real room echo and noise), float16
bolo-tiny.zip	Bolo Tiny: every seed and ablation checkpoint from results/text/, plus the 0.8 MB int8 ONNX. The default is phono_views_aug_asr-s0
bolo-whisper-fp16.zip	Bolo Sun v0 (9 commands, no room/noise augmentation)
bolo-whisper-v1e10-fp16.zip	Bolo Sun v1 trained for 10 epochs: better on synthetic speech, worse on real speech. Rejected; kept for reproducibility
bolo-whisper-wise050-fp16.zip	50/50 weight blend of Swift and Bolo Sun v0: forgets less, learns less

Not included: checkpoints trained on Qwen2.5-3B paraphrases (Qwen Research License), the MuRIL and Qwen baselines (large, and not used by Bolo), and any audio. No recordings of real people saying commands were used in training; the author's own test clips stay on the author's laptop.

Base models: Oriserve/Whisper-Hindi2Hinglish-Swift (Apache-2.0). Training audio was synthesised with ai4bharat/indic-parler-tts (Apache-2.0); room impulse responses and noise from OpenSLR 28.

Honest limits: real-speech accuracy (80% correct actions) is measured on 55 clips from one speaker. See docs/WHAT_WE_TRIED.md.