nakasyou/saikyo-ime-reranker

Saikyo IME Japanese conversion reranker

This repository contains the latest completed LoRA adapter from the saikyo-ime ambiguity training run (vast-experiment-ambiguity-v3/full, 50,000 steps). It ranks Japanese conversion candidates given preceding context and input reading. It is an adapter, not a standalone model: load it with google/t5gemma-2-270m-270m and the project's custom text-encoder score head.

Architecture and input

• Base: google/t5gemma-2-270m-270m (text encoder only)
• PEFT LoRA rank 16, alpha 32; custom scalar score head
• Input: <ctx>...<input>...<output>...<score>
• The included tokenizer defines the four structure tokens.
• Candidate scores are compared within each conversion group.

A generic AutoModelForSequenceClassification loader is not sufficient for this custom architecture. The saikyo-ime T5GemmaReranker implementation defines the score head and preprocessing.

Training

The adapter was continued from a pilot checkpoint and trained on nakasyou/japanese-conversion with an ambiguous-example replay ratio of 0.4. The final run used two epochs, a learning rate of 5e-5, maximum length 256, and seed 20260811. See dataset_manifest.json for the recorded data and input format. The dataset revision was not pinned in the training manifest.

Evaluation

The saved metrics report top-1 accuracy of 0.8516 on the general evaluation set and 0.6719 on the ambiguity evaluation set (256 groups each). These are results from the training run, not an independent benchmark. Full metrics are in eval_metrics.json.

Files

• adapter_model.safetensors, adapter_config.json: LoRA adapter and score head
• tokenizer.json, tokenizer_config.json: tokenizer with structure tokens
• dataset_manifest.json, eval_metrics.json: training provenance and recorded evaluation

The base model is subject to the Gemma license.