sach3v/Gemma_4_E2B_Vision_FOR_Oral_Cancer

Oral Gemma Fine-Tuning Dataset

This repository contains a portable, instruction-tuning dataset for cropped oral mucosal lesion screening. It is designed for vision-language fine-tuning of Gemma-style models on a binary screening task.

Overview

Each example pairs:

• one cropped oral mucosal image
• one short instruction
• one JSON answer with a conservative screening recommendation

This is a screening support dataset, not a diagnostic dataset.

Target labels:

• low_risk_or_variation
• refer_for_clinical_review

The answer format is intentionally strict so the model learns to emit structured, safe JSON.

Dataset Size

Total records: 364

Split summary:

• train: 244 records
  • low_risk_or_variation: 142
  • refer_for_clinical_review: 102
• val: 55 records
  • low_risk_or_variation: 37
  • refer_for_clinical_review: 18
• test: 65 records
  • low_risk_or_variation: 22
  • refer_for_clinical_review: 43

File Layout

• data/train.jsonl
• data/val.jsonl
• data/test.jsonl
• images/train/...
• images/val/...
• images/test/...

Record Format

Each JSONL row contains:

• image: relative path to the packaged image file
• messages: chat-style training example with user and assistant turns
• metadata: split, source class, binary category, crop path, and detector confidence

The image paths are relative so the dataset can be moved between machines without preserving the original local directory structure.

Label Semantics

• source_class preserves the upstream subtype, such as opmd or variation_from_normal.
• binary_category is the actual training target.
• The task framing is conservative and explicitly non-diagnostic.

Expected Task

Input:

• an oral mucosal crop
• a prompt asking for a JSON-only screening response

Output:

• a JSON object containing:
  • category
  • recommendation
  • brief_reason
  • disclaimer

Intended Use

This dataset is intended for:

• fine-tuning a multimodal model on oral screening triage
• research and prototype development
• local/offline screening support workflows

Not intended for:

• standalone diagnosis
• replacing a clinician
• direct clinical decision-making without review

Loading Example

If you want to load the JSONL files locally with Hugging Face datasets:

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files={
        "train": "data/train.jsonl",
        "validation": "data/val.jsonl",
        "test": "data/test.jsonl",
    },
)

Training Notes

• The package creator’s training script uses Unsloth FastVisionModel, UnslothVisionDataCollator, and SFTTrainer.
• The dataset format expects relative image paths inside the repo.
• The current training recipe is conservative and was tuned for 4-bit QLoRA-style fine-tuning.
• An RTX 3090 is sufficient for the current script configuration.

Important Notes

• This repository contains the dataset only. It does not include the trained model weights.
• If you adapt the training script, make sure it does not depend on old absolute paths such as /content/drive/....
• The output should remain conservative: if the image is uncertain or suspicious, the model should prefer refer_for_clinical_review.

Suggested Citation

If you use this dataset, cite it as a cropped oral mucosal lesion screening dataset for binary triage and non-diagnostic review support.