Qwen3-8B-Ja2Zh-SFT (Ollama Version)

This is the Ollama-ready version of the full-parameter merged Qwen3-8B model. It integrates the LoRA weights from enet45/qwen3-8b-ja2zh-lora-v1.1 directly into the base architecture, specifically optimized for Japanese-to-Chinese subtitle translation.

Getting Started

1. Create the Modelfile

Save the following configuration as a file named Modelfile.

[!IMPORTANT] This configuration includes a strict SYSTEM message and a \n stop parameter to ensure the model outputs only the translated line, matching the behavior of the split('\n')[0] logic in the Python script.

# Point to your GGUF file path
FROM ./qwen3-8b-ja2zh-q8_0.gguf

# Matching the LLaMA Factory / Qwen Chat Template
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
"""

# Inference Parameters
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.25
PARAMETER num_ctx 256

# Stop Sequences: Added "\n" to ensure single-line subtitle output
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER stop "\n"

# The same instruction used in the Python SFT script
SYSTEM "将以下日语字幕翻译成中文，文本可能存在识别错误，请根据语境修正后翻译。只输出中文。"

ollama create qwen3-8b-ja2zh-v1 -f Modelfile
ollama run qwen3-8b-ja2zh-v1

Model Features

Context-Aware Correction: Trained to handle typos typical in subtitle OCR.
NSFW Content Support: Capable of handling mature/uncensored content common in various media types.
Efficiency: Optimized via LoRA and merged for zero-latency switching in Ollama.

enet45/qwen3-8b-ja2zh-gguf-v1.1

Qwen3-8B-Ja2Zh-SFT (Ollama Version)

Getting Started

1. Create the Modelfile

Model Features