GeoPOI

The complete trained model includes the full Qwen3-ASR backbone, the full regional parameter bank, Reader, Span, Retriever, reviewed Ranker, six decision heads, and the trained single-head Near/Far/NONE audio Router. Only the public POI directory is restricted to 500 entities.

The regional bank uses one shared right basis (rank 4, with the first 2 rows used by the low-rank tier). This package stores that basis once while retaining every regional core and left-factor coefficient for all 90,760 cells. This is a lossless storage consolidation of the trained V8 checkpoint; regional selection and inference operations are preserved.

from geopoi import GeoPOI
model = GeoPOI.from_pretrained('./GeoPOI', device='cuda')
result = model.transcribe('audio.wav', poi_id='poi_0001')
print(result['text'])

Install requirements.txt; provide a mono16kHz WAV. A public POI ID selects observed context, not a transcript or forced recognition target. Alternatively provide context={'region_id': ..., 'neighborhood_region_ids': [...]}. The audio Router chooses the recall scope; the original decision policy controls later regional intervention and literal rendering.

The offline backend uses real causal audio prefixes with Transformers, preserving first-input-off and later-input Gate permissions. It recomputes prefixes rather than using the production vLLM streaming cache; production latency and benchmark results are not claimed for this backend. Model weights and conditional modules are all loaded; a rejected decision legitimately bypasses intervention.