nowdoor/gemma4-inspect-12b-v1
nowdoor/gemma4-inspect-12b-v1 is a 12-billion parameter language model, fine-tuned from Google's Gemma-4-12B-it, specifically for Korean facility inspection and report review tasks. It excels at comparing inspection results, reviewing defect changes, verifying judgment bases, and assisting with report generation. With a maximum context length of 262,144 tokens, this model is optimized for detailed analysis of technical documents in Korean.
Loading preview...
Model Overview
nowdoor/gemma4-inspect-12b-v1 is a 12-billion parameter language model developed by nowdoor, based on Google's gemma-4-12B-it. This model has undergone full supervised fine-tuning (SFT) to specialize in Korean facility inspection and report review workflows. It supports a substantial maximum context length of 262,144 tokens, making it suitable for processing extensive technical documents.
Key Capabilities
- Korean Q&A: Provides answers regarding facility inspection standards and procedures.
- Comparative Review: Assists in comparing previous and current inspection results for defects.
- Report Drafting: Helps in drafting reports concerning defects, repairs, reinforcements, and follow-up actions.
- Evidence Verification: Identifies missing judgment bases in reports.
- Research Support: Useful for research, experiments, and internal evaluations in the inspection domain.
Observed Strengths
- Quickly identifies numerical errors and inconsistencies between tables and text in reports.
- Structures Korean answers for guidelines, evidence, and improvement recommendations.
- Successfully terminates generation with EOS tokens when output budget is sufficient.
Limitations and Important Considerations
While effective for detecting potential errors and assisting with draft creation, this model is not intended for automated final judgments or independent calculation verification. It may generate factually incorrect information, non-existent standards, or inaccurate figures. Its reliability for detailed arithmetic calculations and final judgment corrections (e.g., "partially adequate") is low. Outputs, especially for critical decisions like safety ratings or budget allocations, must be reviewed by human experts.
When to Use This Model
This model is ideal for supporting human experts in the initial stages of facility inspection report analysis, such as identifying discrepancies or drafting preliminary findings. It should not be used for automated decision-making that directly impacts life or property, or as the sole basis for legal or regulatory judgments.