HIFreee/orena-focus-segment-qwen36-27b-ck1719

Cross-domain temporal adaptation of Qwen3.6-27B for ORena FOCUS SEGMENT

This repository releases the CK1719 model adapter, inference pipeline, training configuration, container recipe, and team-created annotations associated with our MICCAI 2026 ORena SAVE FOCUS SEGMENT submission.

CK1719 is a LoRA adapter for the open-weight `Qwen/Qwen3.6-27B` multimodal model. The upstream base weights are not duplicated here. They can be obtained from the Apache-2.0-licensed Qwen repository and combined with this adapter using PEFT.

Method at a glance

The method addresses three coupled requirements of surgical video question answering: cross-domain adaptation, temporal evidence coverage, and valid response generation.

Cross-domain adaptation without visual forgetting

We train rank-32 LoRA modules on the language backbone while freezing the vision encoder and multimodal projector. This design introduces task-specific semantic and temporal knowledge while preserving the pretrained visual interface. Training uses 13,746 challenge-provided examples from two procedure domains: 8,000 HeiCo-FOCUS examples from colorectal surgery and 5,746 LapChole-FOCUS examples from laparoscopic cholecystectomy.

The two-domain distribution contains natural variation in anatomical context, viewpoint, illumination, reflection, occlusion, and foreign-object appearance. No synthetic low-light, flash, blur, or noise augmentation was used. Released validation and test answers were not included as training examples.

Temporally grounded video representation

The inference pipeline samples the complete video interval at 1 FPS rather than restricting evidence to a single local region. Timestamp-overlayed video is used when available, together with the request's absolute temporal bounds. This design provides uniform sparse temporal coverage and preserves the relationship between visual evidence and the original procedure timeline.

Format-constrained deterministic decoding

A lightweight response controller assigns an output constraint for binary, integer, percentage, foreign-object class, timestamp, multiple-choice, or open-ended questions. Generation is deterministic and limited to 64 new tokens. Post-processing removes only reasoning wrappers and generic answer prefixes. The pipeline contains no answer lookup, question-ID-specific hardcoding, or access to reference answers.

The complete technical description is available in `METHOD_DESCRIPTION.md`.

Training configuration

• —LoRA rank: 32
• —LoRA alpha: 64
• —LoRA dropout: 0.05
• —Trainable component: language-backbone projection layers
• —Frozen components: vision encoder and multimodal projector
• —Training examples: 13,746
• —Epochs: 1
• —Precision: bfloat16
• —Effective batch size: 8
• —Learning rate: 1e-4 with cosine decay
• —Warm-up: 52 steps
• —Context cutoff: 32,768 tokens
• —Video sampling: 1 FPS, up to 300 frames

The complete machine-readable configuration is provided in `training_config.yaml`.

Files

• —adapter_model.safetensors: CK1719 LoRA weights.
• —adapter_config.json: PEFT configuration for the public base model.
• —inference.py: offline ORena FOCUS inference entry point.
• —training_config.yaml: complete training configuration.
• —Dockerfile and requirements.txt: reproducible container recipe.
• —METHOD_DESCRIPTION.md: detailed method and validation disclosure.
• —annotations/cholec80_event_annotations.csv: 12 team-created event labels.

Loading the adapter

import torch
from peft import PeftModel
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration

base_id = "Qwen/Qwen3.6-27B"
adapter_id = "HIFreee/orena-focus-segment-qwen36-27b-ck1719"

processor = AutoProcessor.from_pretrained(base_id)
base = Qwen3_5ForConditionalGeneration.from_pretrained(
    base_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, adapter_id)
model.eval()

For the exact ORena container behavior, use inference.py, the pinned dependencies, and the paths described by the official submission template.

Evaluation disclosure

All reported results were obtained locally on released data. They are not hidden test results or estimates of final leaderboard performance.

The released evaluations completed without inference failures. Peak observed GPU memory remained below 57 GB, and maximum measured per-question latency was 13.91 s on our evaluation server.

Data and annotation terms

The repository does not redistribute ORena challenge videos, LapChole videos, official QA annotations, or hidden evaluation material. Obtain challenge data from the organizers and accept its applicable data-use agreement.

The additional CSV contains semantic labels and clip identifiers or timestamps for 12 clips derived from public Cholec80 material. It contains no video frames or patient metadata. These annotations were used only for diagnostic evaluation and were not used to train CK1719. Source videos remain governed by their original terms.

The Qwen3.6-27B base model is Apache-2.0 licensed. Repository code is released under Apache-2.0. Model-adapter and annotation use additionally remains subject to the upstream model terms, source-data terms, and ORena SAVE FOCUS challenge data agreement; see NOTICE.md and annotations/README.md.

Reproducibility and deployment

The final submission container bundles the base-model snapshot and adapter. It was validated with networking disabled, a single GPU, read-only input, writable output, and the non-root ubuntu user. The functional smoke batch processed all three supplied questions and produced the required answer.json response file.

Limitations

Uniform 1 FPS sampling can miss brief events, particularly under occlusion or rapid camera motion. Long video segments also incur greater computational cost than the released smoke batch. The model can confuse visually similar foreign objects or return inaccurate temporal estimates. It is research software, not a medical device, and must not be used for clinical decision-making.

Acknowledgements

This work uses Qwen3.6-27B and the datasets/tooling provided for the ORena SAVE FOCUS Challenge at MICCAI 2026. Please cite the base model, the ORena FOCUS challenge/datasets, and the underlying surgical datasets as required by their respective documentation.