puwaer/Doujinshi-14b-instruct

TEXT GENERATIONPricing:Input $0.48 / Output $0.96Concurrent Unit Cost:1Model Size:14BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 6, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The puwaer/Doujinshi-14b-instruct is a 14 billion parameter instruction-tuned large language model based on Qwen/Qwen3-14B, specifically fine-tuned for R18 content generation. It was trained using 4 billion tokens of R18-specific data scraped from dmm.co.jp and dlsite.com, making it specialized for information provision and text generation in this domain. This model is designed to follow user instructions for generating R18-related text.

Loading preview...

Doujinshi-14b-instruct: R18-Specialized Instruction Model

The puwaer/Doujinshi-14b-instruct is a 14 billion parameter large language model (LLM) built upon the Qwen/Qwen3-14B architecture. It has undergone continuous pre-training, DPO (Direct Preference Optimization), and SFT (Supervised Fine-Tuning) specifically for R18 content.

Key Characteristics

  • R18 Specialization: Trained on 4 billion tokens of R18-specific data scraped from dmm.co.jp and dlsite.com.
  • Instruction-Tuned: Optimized to follow user instructions for generating text, focusing on information provision and descriptive tasks.
  • Base Model: Utilizes the robust Qwen3-14B as its foundation.

Model Variants

This model is part of a series, each with a distinct focus:

  • Doujinshi-14b-chat: Specialized for natural, everyday conversations.
  • Doujinshi-14b-instruct: Focused on information provision, question answering, and generating text based on specific instructions.
  • Doujinshi-14b-roleplay: Designed for immersive role-playing, maintaining character persona and tone.

Training Data

The model was trained using a combination of datasets, including:

  • puwaer/dlsite-jp-v1, v2, v3
  • puwaer/dmm-fanza-jp-v1, v2, v3
  • puwaer/Doujinshi-sft-dataset-v1
  • puwaer/Doujinshi-dpo-dataset-v1

Licensing

The model is provided under the Apache 2.0 License.