genshiai-daichi/med-lfm2.5-1.2b-autocomplete

TEXT GENERATIONConcurrency Cost:1Model Size:1.2BQuant:BF16Ctx Length:32kPublished:Jun 14, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

The genshiai-daichi/med-lfm2.5-1.2b-autocomplete is a 1.2 billion parameter Japanese medical autocomplete language model, based on the LiquidAI/LFM2.5-1.2B-Base architecture. It has undergone a multi-phase supervised fine-tuning pipeline, specifically adapted for medical text completion and question-answering. This model excels at generating continuations for medical sentences and providing concise answers, making it suitable for applications requiring Japanese medical text auto-completion with a context length of 32768 tokens.

Loading preview...

Overview

This model, med-lfm2.5-1.2b-autocomplete, is a 1.2 billion parameter Japanese medical autocomplete language model developed by genshiai-daichi. It is built upon the LiquidAI/LFM2.5-1.2B-Base architecture and has been progressively fine-tuned through a specialized pipeline to adapt it for medical applications.

Key Capabilities

  • Medical Autocompletion: Optimized to complete Japanese medical sentences, learning to generate continuations given a prefix.
  • Medical QA Adaptation: Incorporates a QA-Completion phase where it learns from medical Q&A datasets, specifically focusing on answer generation to avoid becoming a full Q&A bot.
  • Instruction Following (Minimal): A final QA-SFT phase adds minimal instruction-following capabilities while preserving its primary autocomplete function.
  • Robust Training: Utilizes bf16 precision for stable training and includes pre-processing steps like NFKC normalization and removal of OCR-derived whitespace.

Good For

This model is ideal for use cases requiring Japanese medical text auto-completion. Its phased training ensures it can provide relevant and contextually appropriate continuations for medical phrases and sentences, with a foundational understanding derived from medical Q&A data. Developers can access different training phases (cpt, comp, qa-comp, qa) via revision tags in transformers.