ikedachin/Qwen3.5-9B-base-imabari-CPT-16bit-1e-v1

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 21, 2026License:cc-by-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

ikedachin/Qwen3.5-9B-base-imabari-CPT-16bit-1e-v1 is a 9-billion-parameter causal language model, continually pre-trained by ikedachin on a Japanese Wikipedia corpus focused on the Imabari dialect and Ehime prefecture. Based on Qwen/Qwen3.5-9B-Base, it retains its vision encoder and 32768-token context length, while adapting its language capabilities to a specific Japanese dialect. This model is intended for research into Japanese dialect adaptation and as a base for further fine-tuning on dialect-specific tasks.

Loading preview...

Model Overview

This model, ikedachin/Qwen3.5-9B-base-imabari-CPT-16bit-1e-v1, is a 9-billion-parameter causal language model that has undergone continual pre-training (CPT) on a specialized Japanese Wikipedia corpus. It is built upon the Qwen/Qwen3.5-9B-Base architecture, which features a vision encoder, a hybrid Gated Delta Network and sparse Mixture-of-Experts design for high throughput, and native support for context lengths up to 262k tokens (extensible to over one million).

Key Differentiators & Capabilities

  • Dialect Adaptation: Specifically adapted to the Japanese Imabari dialect and content related to Ehime prefecture through CPT on a ~49 MB corpus from Japanese Wikipedia.
  • Base Model Strengths: Inherits the robust capabilities of Qwen3.5-9B-Base, including its vision-language training, scalable reinforcement learning, and broad multilingual support for over 201 languages and dialects.
  • Research Focus: Primarily intended for research into dialect adaptation and as a foundation for further fine-tuning, rather than a fully aligned chat assistant.

Intended Use Cases

  • Generating or paraphrasing text in the Imabari dialect.
  • Studying continual pre-training strategies on domain-specific or dialect-specific corpora.
  • Serving as a base model for subsequent instruction tuning or RLHF (Reinforcement Learning from Human Feedback).

Limitations

  • The model's general Japanese language capabilities largely depend on the base Qwen3.5-9B model, with only slight improvements in Imabari dialect familiarity.
  • Continual pre-training on Japanese text may slightly reduce performance on other languages.
  • As with all LLMs, it may generate plausible but incorrect statements and is not aligned to avoid harmful outputs without further safety measures.

Finetunes
1 models