OralDetect Finetune skill
A skill for finetuning the OralDetect detector on your dental images: data checks, training launch and a before and after report
- 110 ★ on the repository
- Last commit
Extended access
Why this level
- The skill runs Python scripts and long GPU training on your machine
- Third-party weights and datasets are downloaded from Hugging Face after access approval
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/oraldetect-finetuneDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The skill runs Python scripts and long GPU training on your machine
- Third-party weights and datasets are downloaded from Hugging Face after access approval
This is third-party code. Review the repository files before installing.
What it does
The oraldetect-finetune skill lives in the OralGPT repository, a family of models and benchmarks for digital dentistry. OralDetect detects objects in images of five modalities, and the class is set by its name in a vocabulary, so the model can find things it was never trained on. The skill leads an agent from a folder of your images to a finetuned model in six steps, and every step ends with a checkpoint: the next one does not start until it holds. The steps are: convert labels to COCO, locate the weights, validate the dataset and class vocabulary with a visual check of boxes, build the own_finetune.yaml and own_eval.yaml launchers, score the baseline model, dry-run and start training, then measure the result and compare before and after. The skill's scripts detect the label format (COCO, YOLO, VOC, LabelMe), convert it, validate the dataset, build notebooks with charts and assemble a final HTML page in a Russian or English version. Every script writes a report even on failure, so the reason for a stop is not lost. Precision, recall and F1 before and after sit in one text file, and a figure compares ground truth, the before answer and the after answer.
Who it is for. Computer vision researchers and engineers who finetune a detector on their own dental images and want an agent to handle the routine.
Good fit when
- You have a folder of dental images with labels and need to finetune OralDetect on this data
- Labels come in YOLO, VOC or LabelMe format and must be converted to COCO with a train and validation split
- Before a GPU run you want a dry run of the configuration that catches path errors
- You need a report with metrics and a before and after comparison figure
Not a fit when
- You only need to label images without training: the skill is built for the full finetuning cycle
- Labels are image-level only with no boxes: the skill rejects such data
- You have no GPU or CUDA 12.x environment: training will not start
- You need a clinical conclusion: the repository describes research models and benchmarks
Example request
Finetune OralDetect on my panoramic images in /data/pano, labels in YOLO, and show the metrics before and after.Limitations
It needs CUDA 12.x and one or more GPUs, Python 3.10, torch 2.5.1, mmcv 2.1.0 and pinned numpy and albumentations versions. The 2.7 GB OralDetect-Family weights are downloaded from Hugging Face after requesting access, and open access to the repository does not remove that step. The skill also requires nbformat, nbclient and ipykernel. GitHub shows no license for the repository. These are research models and must not be treated as medical decision tools.
How to disable. Delete the oraldetect-finetune folder from your agent's skills directory. Files the run created sit in the repository's datas, launch_bash and own_finetune directories; remove them by hand if you do not need them.
FAQ
How do I install the oraldetect-finetune skill?
Clone the OralGPT repository and copy OralDetect/skills/oraldetect-finetune into your agent's skills directory, for example .claude/skills. The repository has no separate install command.
What must I prepare before running it?
Three paths: a clone of the repository with run_finetune.py, a folder of images and a directory for weights. The environment is checked with one command that imports torch, mmcv, mmengine, mmdet, transformers, webdataset and yaml.
Which label formats are supported?
The to_coco.py script detects COCO, YOLO, VOC and LabelMe and converts them to COCO. If the format is unknown, the skill asks you to show one label file and guesses nothing.
What do I get at the end?
Precision, recall and F1 before and after in step6_metrics.txt, the step6_compare.png figure and a report.html page. The page comes in Russian and English versions.
Can I use the result in a clinic?
No, the OralGPT README presents research models and benchmarks. The skill prepares training and a report and does not replace clinical validation.
Discussion
No questions or reviews yet.
Related
SkillsVoxel51 skills for computer vision: import datasets, find duplicates, run inference and evaluate models in FiftyOne
Probabl skills for machine learning experiments: a scikit-learn and skrub pipeline, evaluation, audit and a link to Skore Hub
NVIDIA BioNeMo Agent Toolkit: skills for life sciences
NVIDIA BioNeMo Agent Toolkit
NVIDIA skills for biology and chemistry: the agent predicts protein structures, docks molecules and designs proteins
High-Stakes Analytics Lab
High-Stakes Analytics & Decision Lab
An evidence-gated analytics skill for high-stakes decisions: profiles data, picks a method the evidence supports, can answer with a data request
Скилл Claude Code учит агента библиотеке tangermeme для анализа геномных нейросетей: DeepLIFT/SHAP, мутагенез, дизайн.
Excel analysis skill that profiles a sheet, cleans it, reconciles totals against the sheet's own sums and then releases numbers.
README in short
The README presents OralGPT as a family of dental models: multimodal models, the OralAgent agent, the OralDetect detector and benchmarks. For OralDetect it lists weights on Hugging Face, a training set and a benchmark. A separate README in the OralDetect folder covers the environment, pinned library versions, downloading weights and running training and evaluation.
README badge
Original SKILL.md
--- name: oraldetect-finetune description: This skill should be used when the user wants to finetune the OralDetect model on their own dataset — e.g. "finetune OralDetect on my data", "train the dental detector on my data", "run OralDetect on my images". The user provides their dataset folder; validates the COCO format against the loader's real requirements, places the weights, and generates own_*.yaml launchers that are dry-run verified before any GPU is used. version: 1.0.0 --- # Finetune OralDetect on a user-supplied dataset Six steps, in order. Each ends in a **Checkpoint**. Do not start step N+1 until step N's checkpoint holds. A checkpoint that cannot be met is a stop: report what is missing and what the user has to decide. ## Directories Ask the user for these three: ```bash REPO=<the cloned repo — the directory that holds run_finetune.py> IMAGE_ROOT=<the dataset folder the user gave you> WEIGHTS=<where the weights are, or will go> ``` Derive the rest. Everything after this point is addressed through these names: ```bash SKILL=$REPO/skills/oraldetect-finetune # the scripts this skill runs DATAS=$REPO/datas # generated jsons + vocabulary LAUNCH=$REPO/launch_bash # generated yamls OUT=$REPO/own_finetune # checkpoints, logs, predictions REPORT=$OUT/report # every figure and table mkdir -p "$DATAS" "$LAUNCH" "$REPORT" ``` `IMAGE_ROOT` is both the folder the user handed over and the root that `file_name` in the COCO json resolves against. It is the one path outside `$REPO`, because the user's images stay where they are. ## What lands where `$REPORT` is the deliverable — everything a human reads: ``` $REPORT/ ├── step1_format.txt what format the labels were in, what was converted ├── step3_check.txt the validation report ├── step3_preview.png annotated samples ├── step3_preview_classes.png boxes per class ├── step3_preview.ipynb the notebook behind both figures ├── step6_metrics.txt precision / recall / F1, before vs after ├── step6_compare.png ground truth / before / after