protocolsyncllc/medgap-qwen-3.5-fda

Model Summary

This is a specialized fine-tune of Qwen 3.5 9B, optimized for the analysis and interpretation of FDA Warning Letters. It is designed to identify patterns of non-compliance and map observed violations to specific regulatory clauses (21 CFR Part 820 / ISO 13485).

Intended Use

• Regulatory Mapping: Automatic mapping of observed violations to statutory requirements.
• Gap Analysis: Powers Step 3 of the Medgap remediation pipeline.
• Risk Mitigation: Predicts potential audit findings based on historical enforcement logic.

🧬 Development & Data Lineage (Private)

The following resources are maintained in private repositories to prove lineage and ownership. Unauthorized access is restricted:

• Production Model Repository: protocolsyncllc/medgap-qwen-3.5-9b
• Curated Training Dataset: protocolsyncllc/fda-warning-letters
• Inference GGUF: Qwen3.5-9B_q8.gguf

📈 Production Training Metrics (Receipts)

This model was trained using the Unsloth framework on professional-grade infrastructure.

Metric	Value
Training Hardware	NVIDIA A40 (48GB VRAM)
Dataset Size	728 rows (Instruction-Input-Output triplets)
Total Steps	273
Epochs	3
Trainable Parameters	1,966,080 (0.02% of total)
Final Training Loss	1.068
Total Runtime	5,166 seconds (~1h 26m)

Loss Progress

• Initial Loss: 2.197 (Epoch 0.11)
• Mid-point Loss: 0.959 (Epoch 1.32)
• Final Loss: 0.905 (Epoch 2.97)

🛠 Technical Specifications

• Base Model: unsloth/Qwen3.5-9B-Base
• Fine-tuning Method: LoRA / QLoRA
• Inference Format: 8-bit GGUF (Optimized for llama.cpp)
• Software Stack: Torch 2.10.0+cu128 | CUDA 8.6 | Triton 3.6.0
• Context Window: 131,072 tokens

Deployment

This engine is designed for high-performance private inference. The fine-tuned enforcement engine is deployed on dedicated NVIDIA hardware utilizing a native C++/CUDA backend to ensure maximum data security and low-latency remediation. For technical verification or consulting inquiries, visit medgap.org.