mikuhhn1239/qwen3-8b-novel-base-sft
The mikuhhn1239/qwen3-8b-novel-base-sft is an 8-billion parameter language model, fine-tuned from Qwen3-8B-Instruct, specifically designed for generating Chinese web novel content. Trained on 669 Chinese web novels (approximately 72 million characters) covering romance, BL, urban romance, fantasy, and online gaming genres, it excels at learning narrative styles and character dialogue patterns. This model serves as the foundational base for downstream LoRA adapters in the "All Novel Can Be Galgame" workbench, focusing on tasks like narrative parsing, scene segmentation, and character attribution assistance.
Loading preview...
Model Overview
The mikuhhn1239/qwen3-8b-novel-base-sft is an 8-billion parameter model, fine-tuned from the Qwen3-8B-Instruct architecture. Its primary purpose is to serve as a base model for generating and processing Chinese web novel content, specifically within the "All Novel Can Be Galgame" workbench.
Key Capabilities & Training
This model was trained using 669 Chinese web novels, totaling approximately 72 million characters. The training data encompasses popular genres such as romance, BL (Boys' Love), urban romance, fantasy, and online gaming. The fine-tuning process involved a full-parameter SFT (Supervised Fine-Tuning) approach, utilizing a combined dataset of 72,573 entries for continuation and instruction-based writing. Training was conducted on 4 A800 80GB GPUs with DeepSpeed ZeRO-2, achieving a final loss of 2.47 over 2 epochs with a maximum sequence length of 2048.
Intended Use Cases
This model is specifically designed to learn and replicate the narrative styles and character dialogue patterns prevalent in Chinese web novels. It acts as a foundational base for specialized LoRA adapters, enabling advanced tasks such as:
- Narrative Unit Classification: Identifying and categorizing different narrative elements.
- Scene Boundary Detection: Pinpointing the start and end points of scenes within a text.
- Character Attribution Assistance: Helping to correctly assign dialogue and actions to specific characters.
Limitations
It is important to note that this model is optimized for Chinese input only and is not a general-purpose instruction model. Its training data is exclusively focused on web novels, and it lacks safety alignment, making it unsuitable for generating sensitive content.