Solo project · ORena SAVE FOCUS 2026
I adapted Qwen3-VL-4B to surgical foreign-object question answering, then built a temporal search policy that spends frames where the question needs them. The result is two complementary, fully local inference systems.
“When is the needle
first seen?”
00:33:00Illustrative answerThe project
Short segments need precise local evidence; hour-long procedures need a search strategy. Both repositories share the surgical adaptation recipe, answer-format constraints, and absolute-time frame overlays.
01 / DISCOVR-SEGMENT
One merged checkpoint answers segment questions. A predicted timestamp triggers a second look within ±15 seconds; timestamps stated in a question guide ±30-second evidence windows.
02 / DISCOVR-PROCEDURE
W64 supplies a broad event pointer. NW2 refines it twice and answers other questions. Models load sequentially to fit the validation GPU; needle reappearance uses four post-anchor presence checks.
01 / Post-training
I constructed surgical scene-literacy data from SSG-VQA and CholecT45, continued the adapters on joint FRAME, SEGMENT, and PROCEDURE supervision, and merged each selected adapter into its own bfloat16 model.
Join surgical QA to source frames, de-duplicate records, normalize formats, and represent each example as image-conditioned chat.
Apply supervised cross-entropy to answer tokens, masking the system instruction, question, padding, and visual placeholders.
Use chronological frames, a 768-pixel longest-side cap, and absolute video timestamps throughout temporal training and inference.
Training loss was monitored with row-random splits, so it is not a video-disjoint generalization estimate. This project uses supervised fine-tuning; the temporal search policy is deterministic.
02 / Inference
A uniform pass can miss a brief event in a long video. I used the first prediction as a pointer, then resampled near it. Narrowing the interval improves temporal resolution with a bounded frame budget.
W64 scans the complete procedure; NW2 then refines inside 1,200-second and 100-second windows. Only one model is resident at a time.
The search uses predicted timestamps to select evidence. Gold timestamps are used for evaluation, and the selected frames always keep the absolute video clock.
03 / Experiment design
These ablations ask whether changing evidence placement improves timestamp answers. Each row is a separate local study with its own question subset and checkpoint context.
| Intervention | Local accuracy | Questions | Scope |
|---|---|---|---|
| Segment · predicted-time refinement | 50.4% → 55.0% | 1,577 | Earlier Segment checkpoint; single-timestamp answers |
| Segment · question-anchored windows | 72.7% → 79.1% | 422 | Questions stating a time; same 64-frame budget |
| Procedure · three-pass search | 6.8% → 35.1% | 746 | Single-checkpoint study; single-timestamp answers |
These subset results are not overall challenge scores, and the gains should not be added. The recorded challenge pre-evaluation values are shown separately below.
Challenge pre-evaluation · recorded in the released provenance
04 / Engineering
I released the merged checkpoints, inference code, method descriptions, and container tooling. The public source can be reviewed and tested on CPU before running a full GPU smoke test.
Each track remains independently buildable. This page connects the common design and the evidence behind both releases.
Segment source + method ↗ Procedure source + method ↗ Three-slide project overview / PDF Project slides / PowerPoint Figure generation sourceThis is a research challenge release. Dataset access remains subject to the original terms; patient videos and raw challenge annotations are not distributed. The repositories provide inference artifacts and documented training recipes; complete training scripts and manifests are not included.
