Grayye233/BN4601A-Colonoscopy-Models-Handoff

BN4601A complete models and English Colab handoff

Download and start

Download the complete English package (1.79 GB compressed; 2.24 GB extracted).

Open the download/setup notebook in Colab. Run it to place the verified complete folder in your own Google Drive. Then open notebook 02 for core inference, 05 for direct VQA, or 03 for future Creative Work LoRA.

The archive contains full model directories and processor/tokenizer files, not adapters. Extract it before loading models; the repository root itself is not an AutoModel checkpoint.

Included checkpoint	Parameters	Weight file size	Purpose
artifacts/vqa_vlm/model	507,482,304	1.015 GB	Selected stage-2 VQA continuation; use this as the future LoRA base
artifacts/base_vlm/model	507,482,304	1.015 GB	Preserved stage-1 domain checkpoint and fallback
artifacts/segmentation/best.pt	14,328,081	57.4 MB	ResNet18 U-Net segmentation

Both VLM checkpoints use complete full-weight training; no Creative Work LoRA has been trained. BF16 weights use approximately two bytes per parameter. File size is not GPU memory usage: training additionally needs activations, gradients and optimizer state.

Included: five original Colab notebooks, code, checkpoints, processors, measured evaluation outputs, selected VQA examples and source notices. The CVC-containing selected-case illustration and all 612 CVC predicted-mask PNGs are removed from the public archive. Numerical CVC results remain; no CVC source image bytes are distributed. All complete checkpoint files are retained.

Results and remaining work

Stage-2 original-question source-label agreement: 66.2%; reserved paraphrase: 37.0%. Core input-evidence field agreement: 99.54%. These are bounded research benchmarks, not clinical accuracy or validated fluency. Read VQA_UPDATE.md, including failure cases and limitations. Creative Work dialogue review, LoRA, interface, final report and slides remain next-team tasks. Actual training and evaluation were local; real Colab GPU execution and the new setup notebook have not been exercised in Colab.

Shared data: BN4601A-Colonoscopy-Curated, configurations segmentation, core_sft and vqa_sft.

Attribution and integrity

Read SOURCES.md, MODEL_CARD.md and licenses/. Upstream SmolVLM is Apache 2.0; this does not override source dataset noncommercial restrictions. Preserve all source attribution. Intended for academic research and education, not clinical deployment.

Archive SHA-256: baf8c5608f5a245fde89daec47e7c639b10f93cf37ab7b569c3f4127689b3ffd. The extracted package includes an updated INTEGRITY_SHA256.json and verify_package.py. The original local course handoff is preserved separately.