katsukiono/gemma3-270m-pred-dpo

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.3BQuant:BF16Context Size:32kPublished:Jan 3, 2026Architecture:Transformer Featherless Exclusive Cold

The katsukiono/gemma3-270m-pred-dpo is a 270 million parameter Gemma 3 model, fine-tuned by katsukiono using SFT and DPO methods. Optimized specifically for Japanese IME (Input Method Editor) predictive text, it generates word predictions following a `[---]` boundary within input text. This lightweight model is designed for efficient, real-time text prediction on devices like iOS, offering fast token generation rates.

Loading preview...

Model Overview

katsukiono/gemma3-270m-pred-dpo is a lightweight 270 million parameter model based on google/gemma-3-270m-it, specifically optimized for Japanese IME (Input Method Editor) predictive text. It has undergone Supervised Fine-Tuning (SFT) followed by Direct Preference Optimization (DPO) to enhance its prediction capabilities.

Key Capabilities

  • Predictive Text Generation: Designed to predict words that follow a [---] boundary in user input, simulating a keyboard's predictive text feature.
  • Lightweight and Efficient: At 270M parameters, it's suitable for deployment on resource-constrained devices, demonstrated by iOS performance metrics of 12 ms/token and 82.94 tokens/sec.
  • Japanese Language Focus: Optimized for Japanese text prediction, making it ideal for applications requiring high-quality Japanese IME functionality.
  • Flexible Deployment: Distributed in both Transformers (Hugging Face) format and GGUF (f16 / Q4_K_M) for compatibility with various inference frameworks like llama.cpp.

Usage and Integration

Users must adhere to a specific input format, including a fixed prompt and the [---] boundary to delineate user input from the prediction target. The model is designed to output a single line containing the [---] boundary and predicted words, which can then be split by / to extract candidate suggestions. Post-processing recommendations include adopting only the first line of output and checking for the presence of [---] for robust production use.

Licensing

This model inherits the usage terms and license of its base model, google/gemma-3-270m-it.

Finetunes
1 models