Qwen/Qwen3.8-27B
Image-Text-to-Text • 28B • Updated • 2.09M • • 12.1k
Brad Hayes
brad
·
AI & ML interests
Explainable AI, Human-Robot Interaction, Human-Machine Teaming, Agents, VLM, VLA, Agent-Human Task Handover
Recent Activity
Organizations
Add the B2 human review of the drone nominal pool Frames leave the pool for three independent reasons, reported separately and never summed: hazard 4/427 (0.9%), nadir framing 276, ground-level capture 29. Usable pool 120. Anything computing a deployment false-alarm rate against the scene inventories must apply exclude_ids_combined — without it the rate is scored over 307 frames a human already rejected.
#9 opened 2 days ago
by 
brad
Add Stage-1 scene inventories for the drone nominal pool (427 frames) AeroScapes + Wilds Drones, Stage-1 only by contract: attaching hazard text to scenes whose entire value is carrying none would make any false-alarm rate computed against them circular. Pair with reviews/b2_contamination_result.json, which cuts the usable pool from 427 to 120.
#8 opened 2 days ago
by brad
Add neutral scene grounding for the dense corpus (2,018 frames) Stage-1 grounding produced by qwen3.6-27b: what is visible, no safety judgment. This is the evidence the deployed SOS pipeline scores, as opposed to the hazard annotations, which were written by a hazard-seeking prompt. Reasoning mode OFF — with it on, the model's untagged chain-of-thought is parsed as scene sentences (56-58 per frame instead of 8-10, carrying the very hazard words the grounding rule forbids). D4 lint: 3 warnings / 2,018 frames; 0 overlong.
#7 opened 2 days ago
by brad
Upload dataset
#3 opened 6 days ago
by brad
Upload dataset
#2 opened 7 days ago
by brad
liked a
model 8 days ago
updated a
model about 4 years ago
updated 2
models over 4 years ago