mikuhhn1239/qwen3-8b-novel-base-sft

Model Overview

This model, mikuhhn1239/qwen3-8b-novel-base-sft, is a full-parameter fine-tuned version of Qwen3-8B. It was trained on a dataset comprising 669 Chinese web novels, totaling approximately 72 million characters, across genres like romance, BL, urban romance, fantasy, and online gaming. The primary goal of this training was to enable the model to learn the distinctive narrative styles and character dialogue patterns prevalent in Chinese web novels.

Key Capabilities & Purpose

This model functions as the Stage1 base model for the "All Novel Can Be Galgame" workbench. It is designed to serve as a foundational layer for subsequent LoRA adapters, which perform specific agent tasks:

• Narrative Unit Classification: Via mikuhhn1239/qwen3-8b-narrative-parsing-lora
• Scene Boundary Detection: Via mikuhhn1239/qwen3-8b-scene-segmentation-lora
• Character Attribution Assistance: Via mikuhhn1239/qwen3-8b-attribution-assist-lora

Training Details

The model underwent full-parameter Supervised Fine-Tuning (SFT) using a dataset of 72,573 entries, formatted in ChatML ([system, user, assistant]). The training involved a maximum sequence length of 2048, 2 epochs, and an effective batch size of 64. It utilized DeepSpeed ZeRO-2 on 4 A800 80GB GPUs, achieving a loss reduction from 3.36 to 2.47 over approximately 9 hours.

Limitations

• Language Support: Primarily supports Chinese input.
• Domain Specificity: Optimized for web novel content; not a general-purpose instruction model.
• Safety: Lacks safety alignment, making it unsuitable for sensitive content generation.