jtmaxsoft/OFKMS-Migration-Qwen3.5-9B-SFT
The jtmaxsoft/OFKMS-Migration-Qwen3.5-9B-SFT model is an 8 billion parameter Qwen3.5-based language model developed by TmaxSoft Japan, fine-tuned for mainframe migration design tasks. It specializes in converting COBOL, JCL, and Assembler code, leveraging a proprietary migration knowledge base. This model excels at generating migration designs and diagnosing error patterns for systems transitioning to OpenFrame environments. With a 32768 token context length, it provides comprehensive support for complex migration projects.
Loading preview...
OFKMS-Migration-Qwen3.5-9B-SFT Overview
This model, developed by TmaxSoft Japan, is a specialized Qwen3.5-9B variant fine-tuned for mainframe migration design. It leverages QLoRA (SFT) with a training dataset comprising 1,288 SFT entries and an equal number of DPO pairs, derived from TmaxSoft Japan's proprietary migration knowledge base. The model supports Japanese, Korean, and English, and is available in a bf16 variant.
Key Capabilities
- COBOL Migration: Analyzes COBOL source patterns and generates conversion rules.
- JCL Transformation: Facilitates migration from standard JCL to OpenFrame JCL.
- Assembler Conversion: Supports migration of Assembler code to C or OFASM.
- Design Document Generation: Aids in creating comprehensive migration design documents.
- Error Diagnosis: Identifies and diagnoses common error patterns, including ABEND codes and JES messages.
Training Details
The model was fine-tuned using QLoRA with a rank of 64 and alpha of 128, resulting in 174 million trainable parameters (2.09% of the total 8.4 billion). Training involved 3 epochs with an effective batch size of 64, utilizing a cosine learning rate schedule of 2e-5 on NVIDIA A100 40GB hardware.
Good For
This model is ideal for developers and organizations involved in modernizing legacy mainframe systems. Its specialized knowledge makes it particularly effective for automating and streamlining the complex process of migrating COBOL, JCL, and Assembler applications to contemporary platforms like OpenFrame.