Medical AI Engineer — Dental SLM Fine-tuning (QLoRA, Vision + Language Models)

Job Title: Medical AI Engineer — Dental SLM Fine-tuning (QLoRA, Vision + Language Models) Overview We are building a proprietary dental aligner AI system for a dental company client. The system uses two models running on-premise: DentalGemma 1.5 4B (vision) — reads intraoral photos and X-rays, outputs structured clinical JSON Qwen3 8B (language) — reads clinical findings and generates aligner treatment plans Both models need fine-tuning on our proprietary dataset of 16 annotated patient cases with before/after photos, X-rays, clinical notes, and doctor-written treatment plans. We have the data ready. We need someone to execute the fine-tuning pipeline. Scope of work Task 1 — DentalGemma fine-tuning Fine-tune DentalGemma 1.5 4B on our dental image dataset using QLoRA Input: intraoral photos + X-rays with annotated JSON labels Output: model that correctly identifies teeth in FDI notation, classifies malocclusion, crowding severity, and aligner suitability Deliver: merged model weights + GGUF Q4_K_M converted file Task 2 — Qwen3 8B fine-tuning Fine-tune Qwen3 8B on our SFT dataset using QLoRA Input: clinical findings JSON + doctor treatment plan pairs Output: model that generates treatment plans matching our doctor's clinical style Deliver: merged model weights + GGUF Q4_K_M converted file Task 3 — Evaluation and benchmarking Run both models on 5 held-out test cases Compare outputs against ground truth Provide accuracy report showing improvement over base models Task 4 — Documentation Clean Python training scripts committed to our private GitHub repo README with exact commands to reproduce training W&B training charts showing loss curves What we provide Private GitHub repo with full codebase 16 annotated patient records (photos, X-rays, treatment plans) HuggingFace account with model access RunPod/Vast.ai credits for GPU compute Clear JSON schema and SFT dataset format Daily availability for questions Required skills HuggingFace transformers, peft, trl, bitsandbytes QLoRA fine-tuning experience on vision-language models Experience with medical or domain-specific model fine-tuning Python, Git GGUF conversion with llama.cpp Nice to have Previous MedGemma or Gemma fine-tuning experience Dental or medical AI background Unsloth experience (faster training) Deliverables Fine-tuned DentalGemma GGUF file Fine-tuned Qwen3 8B GGUF file Evaluation report (5 test cases) Training scripts in GitHub W&B training charts To apply, please answer: Have you fine-tuned a vision-language model before? Which one? Have you used QLoRA with bitsandbytes on a 4B+ parameter model? What is your estimated timeline for Tasks 1 and 2? Share one relevant project from your portfolio