jtmaxsoft/OFKMS-Migration-Qwen3.5-9B-DPO

OFKMS Migration Design - Qwen3.5-9B DPO (bf16)

Mainframe migration design specialized model fine-tuned from Qwen3.5-9B.

Model Description

This model is fine-tuned for COBOL/JCL/Assembler migration design tasks, trained on TmaxSoft Japan's proprietary migration knowledge base.

• Base Model: Qwen3.5-9B
• Fine-tuning: QLoRA (DPO)
• Training Data: 1,288 SFT entries + 1,288 DPO pairs
• Languages: Japanese (primary), Korean, English
• Variant: bf16

Training Details

• Method: QLoRA (rank=64, alpha=128)
• Trainable params: 174M / 8.4B (2.09%)
• Epochs: 3
• Batch size: 4 (gradient accumulation: 16, effective: 64)
• Learning rate: 2e-5 (cosine schedule)
• Hardware: NVIDIA A100 40GB

Supported Tasks

• COBOL source pattern analysis and conversion rules
• JCL to OpenFrame JCL migration
• Assembler to C/OFASM migration
• Migration design document generation
• Error pattern diagnosis (ABEND codes, JES messages)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("jtmaxsoft/OFKMS-Migration-Qwen3.5-9B-DPO")
tokenizer = AutoTokenizer.from_pretrained("jtmaxsoft/OFKMS-Migration-Qwen3.5-9B-DPO")

prompt = "COBOL PERFORM statement OpenFrame migration pattern"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Organization

TmaxSoft Japan