kungfupanda12/bolo-tiny

Bolo Tiny: a 0.8 MB Hinglish command model

Bolo Tiny is the command-understanding model in Bolo, an offline Hinglish voice-command assistant. It reads a command in Roman, Devanagari or mixed script and returns the intent plus slots (folder, app, contact, level, and so on). There are 20 command types plus "none".

"bhai rust learning folder mein terminal kholo"  ->  {intent: open_terminal, folder: "rust learning"}

• Built from scratch: 0.69M parameters. A 2-layer transformer over PhonoHash features: hashed character n-grams of the text, its transliteration, a phonetic key and a consonant skeleton. Spelling variants like kholo / khol do / खोलो therefore land close together.
• Size and speed: 0.8 MB as int8 ONNX, about 1 ms on one CPU thread.

Files

File	What it is
model.safetensors (also model.pt, a plain state dict) + config.json + threshold.json	phono_views_aug_asr-s0, the default in Bolo. Trained on synthetic template text plus transcripts of TTS clips made by the speech model, so it has seen realistic speech-recognition errors
onnx/bolo_tiny.int8.onnx, onnx/bolo_tiny.onnx	ONNX export of the earlier text-only model phono_views_aug-s0 (88.3% exact match, same as PyTorch)

Loading uses the bolo package (bolo.models.tiny) from the GitHub repo, because the PhonoHash featuriser is part of it:

from bolo.pipeline import Bolo
bolo = Bolo.load(nlu_path="path/to/this/folder")
bolo.understand("volume 40 pe kar do")   # {'intent': 'set_volume', 'slots': {'level': 40}, ...}

Results

	Synthetic test exact match	False actions	Size	Latency
Hand-written rules	84.9	18.7%	–	0.01 ms
Bolo Tiny (3 seeds)	87.4	5.2%	0.8 MB	about 1 ms
MuRIL-base fine-tuned	91.6	3.6%	950 MB	15.6 ms

With grounding (checking that the folder or app exists on the machine), Bolo Tiny gets 95.8% vs MuRIL's 96.4%, at about 1/1000 of the size. Real-speech results are on the Bolo Sun card; they come from one speaker so far.

Ablations, including what failed:

• PhonoHash: +5.0
• data augmentation: +6.4
• multi-script views: +2.5
• a cross-script consistency loss: −1.4
• distillation from MuRIL: −4.0

Full details are in WHAT_WE_TRIED.md. Every seed and ablation checkpoint is on the GitHub release.

Known weaknesses:

• Confident mistakes on near-miss phrasings ("naya folder banao").
• "kuch bhi random baat" is read as calling a contact named "random".
• All training text is synthetic.

License

Apache-2.0. Training text came from our own templates; no LLM-generated data is included in this model.