Div97/orena-focus-segment-fullvis-w64

DISCOVR-SEGMENT FullVis-W64

Merged bfloat16 checkpoint used by DISCOVR-SEGMENT, submitted to the ORena SAVE FOCUS 2026 SEGMENT track by team Incision Impossible.

• —Base: Qwen/Qwen3-VL-4B-Instruct
• —Adaptation: language, vision-encoder, and merger LoRA
• —Merge: PEFT merge_and_unload
• —Serving budget: 64 frames
• —Model SHA-256: 622fd66547b2ad88f9fcf9c74a22450f44b4c88cef8fcf1a9b464de2a51dcff3

Source and documentation:

• —GitHub source
• —Concise implementation method
• —Scientific method description
• —Training provenance

Training summary

FullVis-W64 was trained in two stages:

1. 1.one-epoch surgical scene-literacy warm-up on a 238,925-row SSG-VQA / CholecT45 manifest, with language rank 16, vision rank 16, and merger rank 64; and
2. 2.three scheduled epochs of continued all-track FOCUS fine-tuning on 34,290 HeiCo and LapChole training rows with a maximum of 64 frames.

The detailed document presents the motivation, data construction, visual preprocessing, optimization objective, training schedule, temporal inference policy, evaluation, and limitations.

Inference method

DISCOVR-SEGMENT uses question-dependent temporal routing:

• —ordinary questions sample 64 frames across the complete segment;
• —questions containing timestamps use up to two ±30-second evidence windows;
• —single-timestamp questions receive a second 64-frame pass in a 30-second window around the first prediction;
• —temporal frames receive an absolute-procedure-time overlay; and
• —output is normalized to the inferred answer format.

Intended use and limitations

The checkpoint is intended for non-commercial research and challenge reproduction. It is not a medical device and must not be used for clinical decision making. No patient videos or raw challenge annotations are distributed here.

License and data terms

The Qwen3-VL base model and released source code use Apache-2.0. Training also used SSG-VQA, whose repository specifies CC BY-NC-SA 4.0 for non-commercial scientific research, plus challenge datasets governed by their respective owners. The license: other metadata reflects these mixed terms; it does not replace any source-dataset license or access agreement.