OrenaFocusVQA-jmees โ weights
Trained weights for team JMEES's solution to the Orena Focus Challenge (MICCAI 2026), surgical video question answering on HeiCo-FOCUS / LapChole-FOCUS.
Code, training recipes and reproduction instructions: https://github.com/JmeesInc/OrenaFocusVQA-jmees
These files are the contents of each track's container/resources/. The directory names are what the
inference container looks for โ keep them.
Layout
| path | what | size |
|---|---|---|
frame/m1_m03b_overlay |
LoRA adapter, route A (also used as m2_m03b_768 at a different resolution) |
196 MB |
frame/m3_v06_r32 |
LoRA adapter, route A, no overlay | 392 MB |
frame/b1_q02b_r1, b2_q00b_r4, b3_q03b_r1sa |
LoRA adapters, route B (all-data) | 196 MB each |
frame/detector, frame/detector_clip |
Mask2Former foreign-object instance segmentation | 822 MB each |
segment/adapter_n04 |
LoRA adapter, 1st model | 196 MB |
segment/adapter_r00, adapter_r00c |
LoRA adapters, 2nd model (out of / in distribution) | 196 MB each |
procedure/adapter, adapter_video |
LoRA adapters, image pass and video pass | 196 MB each |
procedure/m2f, m2f_sm89.trt |
Mask2Former detector and its TensorRT engine | 822 / 450 MB |
procedure/phase_cholec, phase_heico |
surgical phase index (ConvNeXtV2-tiny + MS-TCN, strides 5/10/20) | ~110 MB each |
*/group_templates.json, time_count_prior.json, m2f_classes.json |
routing and post-processing tables | small |
All LoRA adapters are rank 16 / ฮฑ 32 on Qwen/Qwen3.5-9B, trained in 4-bit NF4 with bf16 compute. They are adapters, not merged models โ load the base model and apply the adapter; merging a LoRA into a 4-bit base silently discards it.
Download
huggingface-cli download negichi/OrenaFocusVQA-jmees-weights --include 'frame/*' \
--local-dir FRAME/container/resources
Each track's container/weights_manifest.json in the GitHub repository lists the md5 of every file,
matching what container/stage_resources.sh asserts, so a download can be verified against the
training run that produced it.
License
The weights are derived from Qwen3.5-9B and trained on challenge data released for research use; they are provided for non-commercial research purposes. The base model and each training corpus remain under their own licenses.