bbanany/qwen25-3b-korean-pii-qlora3

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 9, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

The bbanany/qwen25-3b-korean-pii-qlora3 model is a 3.1 billion parameter Qwen2.5-3B-Instruct base model, fine-tuned using QLoRA for Korean PII (Personally Identifiable Information) classification. It specializes in identifying whether candidate spans within Korean RAG (Retrieval Augmented Generation) responses correspond to PII or NOT_PII. This model achieves high accuracy (0.994453) and macro F1 (0.994299) on its test set, making it suitable for privacy-sensitive Korean text processing.

Loading preview...

Overview

This model, bbanany/qwen25-3b-korean-pii-qlora3, is a specialized fine-tune of the Qwen2.5-3B-Instruct base model, developed by bbanany. It leverages QLoRA (4-bit NF4 with double quantization) to adapt the 3.1 billion parameter model for Korean PII classification. Its primary function is to determine if a given text span within a Korean RAG response contains Personally Identifiable Information (PII) or not (NOT_PII). The model's weights are provided as merged FP16 safetensors, allowing for inference without bitsandbytes or PEFT, and supporting both NVIDIA CUDA and Apple Silicon MPS.

Key Capabilities

  • Korean PII Classification: Specifically trained to classify candidate text spans in Korean RAG responses as PII or NOT_PII.
  • High Performance: Achieved a test accuracy of 0.994453 and a macro F1 score of 0.994299 on its dedicated test set.
  • Efficient Inference: Merged FP16 safetensors enable direct loading and inference on CUDA and MPS devices without additional libraries like bitsandbytes or PEFT.
  • Optimized Training: Fine-tuned using 4-bit NF4 QLoRA on attention and MLP projection layers with a rank/alpha of 16/32, and a maximum sequence length of 512.

Limitations

  • Limited Output Classes: Only classifies into PII and NOT_PII.
  • Domain Specificity: Trained on synthetic/business-context data; requires separate validation for real-world operational documents, new tags, or language variations.
  • Not a Legal Tool: This model is not intended as a legal judgment tool for PII processing.

License

The model weights are subject to the Qwen Research License, primarily for non-commercial research and evaluation. Commercial use may require separate permission from the original licensor. The accompanying code files (inference.py, requirements-inference.txt) are provided under the MIT License.

Finetunes
1 models