fa0311's picture
Publish curated IME context dataset
3baaa4a verified
|
Raw
History Blame Contribute Delete
1.96 kB
metadata
language:
  - ja
license: cc-by-4.0
pretty_name: IME Context Dataset
size_categories:
  - 100K<n<1M
task_categories:
  - text-classification
tags:
  - ime
  - japanese-input-method
configs:
  - config_name: core
    default: true
    data_files:
      - split: train
        path: data/core.jsonl
  - config_name: hard_context
    data_files:
      - split: train
        path: data/hard_context.jsonl
  - config_name: person_context
    data_files:
      - split: train
        path: data/person_context.jsonl
  - config_name: difficult_context
    data_files:
      - split: train
        path: data/difficult_context.jsonl

IME Context Dataset

Canonical minimal dataset for Japanese IME candidate reranking. Every row has only context, reading, and target.

Configurations

  • core: 8,084
  • anonymized person context: 219
  • synthetic hard context: 146,270
  • difficult context: 11,329

person_context replaces concrete people and public persona names with [PERSON]. hard_context is kept separate so callers can choose its sampling weight rather than allowing a large synthetic set to dominate ordinary data. All configurations are training material. Create a deterministic train/validation split when training; use an independent locked benchmark as test data.

Exact rows present in the project's locked evaluation set were excluded from the hard training data. Dictionary candidate lists are not redistributed.

License

This dataset is licensed under the Creative Commons Attribution 4.0 International license (CC-BY-4.0). Generated text was reviewed, normalized, deduplicated, and filtered against the locked evaluation data before release.

Related repositories