BimOmni · Barbados-adapted multimodal model

BimOmni hears Barbados properly.

A small LoRA on top of Qwen3-Omni-30B-A3B-Instruct, trained on ten years of the Barbados Advocate. Published openly on the Hugging Face Hub as a fused bf16 checkpoint and a 4-bit MLX snapshot that runs on a MacBook.

The BimOmni Barbados fine-tune Space on Hugging Face, comparing Whisper-large-v3, Qwen3-Omni (base) and BimOmni fine-tune transcriptions of the same Barbados audio clip.
The Hub Space · Whisper vs Qwen3-Omni vs BimOmni on a Barbados clip.
+13.3 pp
on the 60-probe Barbados knowledge test
10 / 10
Barbados proper nouns, vs 2–3 for the base model
1,624 / 1,624
optimiser steps, on a single rented H200
bf16 + 4-bit MLX
fused snapshot and a MacBook-friendly MLX build

The model

A reproducible template, not a finished product

BimOmni is one part of Pulse, a public-signal intelligence system for Barbados built during the Future Caribbean buildathon. Pulse ingests fragmented social, broadcast and news signals and resolves them into a graph of sources, places, events, people and communities. The island-wide product is live at pulsebarbados.com.

The model itself is not tied to Pulse. It can be used independently for Caribbean-domain transcription, information extraction, search, summarisation, research and other multimodal applications. It is a research artefact built during a 21-day buildathon, not a commercial product — treat its outputs as a starting point for human review rather than as ground truth.

What it is, concretely: a parameter-efficient LoRA on Qwen3-Omni's text-only "Thinker", trained on roughly 51.6 million tokens of cleaned Barbados newspaper text. The audio and visual encoders are frozen. The model is better at choosing between plausible outputs given ambiguous audio — not at hearing the audio itself.

Evaluation

Three tracks, three answers, no theatrical claims

BimOmni was scored on three independent tracks. The headline number is honest and small: it is a research artefact, not AGI, and the eval sets are deliberately tiny. Reproduction details and exact scores live in the model card.

Bar chart: Qwen3-Omni base 61.7%, attention-only LoRA 55.0%, attention+MLP/MoE LoRA 70.0%, full schedule 75.0%.
60-probe Barbados knowledge test, run on the same 4-bit MLX stack. Reproducible to zero percentage points.
Track Qwen3-Omni base BimOmni Note
Barbados knowledge (60 probes) 61.7% 75.0% +13.3 pp overall
Local (20 well-known probes) 55.0% 70.0% +15.0 pp
Rare local (20 niche probes) 40.0% 60.0% +20.0 pp
General controls (20 probes) 90.0% 95.0% +5.0 pp

Proper nouns

The concrete example worth quoting

The clearest signal is the proper-noun test. Whisper transcribed Carlisle Bay as "Carlyle Bay" and Oistins Fish Fry as "Oistin's Fish Fry". BimOmni transcribed all ten local proper nouns correctly, against two or three for the unadapted baselines.

Reference Whisper Qwen3-Omni BimOmni
Worthing Square Wooding Square Willing Square Worthing Square
Chefette Shafet Chefette Chefette
Richard Haynes Richard Hines Richard Hinds Richard Haynes
Q.P. Bistro Kupi Bistro QPB Bistro Q.P. Bistro
Oistins Fish Fry Oistin's Fish Fry Oyston's Fish Fry Oistins Fish Fry
Puddin' and Souse putting on south Puddin' and Sauce Puddin' and Souse
Farley Hill Farley Hill Carlisle Hill Farley Hill
Animal Flower Cave Animal Flower Cave Animaux Flower Cave Animal Flower Cave
Carlisle Bay Carlyle Bay Carlisle Bay Carlisle Bay
Champers Champas Champers Champers

These ten examples were selected to examine local proper-noun handling. They are not a complete word-error-rate benchmark and the two-video comparison should not be read as a general ranking of transcription systems. Whisper remained substantially faster on both samples.

Interview

Voice of Barbados · NewsMakers

Voice of Barbados 92.9 FM
NewsMakers interview with Matt Hamilton · Dharach

In this NewsMakers conversation on Voice of Barbados 92.9 FM I cover BimOmni end to end: why a frontier model still mishears a Bajan accent, what's missing when you ask a generic assistant for a "faucet" in Barbados, how the Barbados Advocate archive became 51.6 million tokens of training data, and what kadooment and cobbler pot taught a model that off-the-shelf speech recognition had never learnt.

Yuh Gettin Tru and WeOutside246 come up as well — they are the same idea in product form. Island-wide product search across 100+ retailers, and an events aggregator that keeps the Crop Over calendar honest. Both exist because generic search drops you the moment local context matters.

I moved to Barbados on the Welcome Stamp in August 2021. BimOmni is what came of living here long enough to know what the assistants were missing.

Listen to the interview

Provenance

The artefacts and where to find them

Everything is published openly. The training corpus is not redistributed — it is drawn from a public Caribbean Digital Library archive, used under research-purpose terms.

Want to ship something real?

I'm available for consulting on civic tech, useful AI, and digital public infrastructure across the Caribbean and beyond.

See consulting