Datasets:

KeisukeMiyamoto
/

lambda-chat

language:
  - ja
pretty_name: lambda-chat
license: other
task_categories:
  - text-generation
tags:
  - japanese
  - instruction-following
  - conversational
  - parquet
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files:
      - split: train
        path: train/*.parquet
      - split: validation
        path: validation.parquet
      - split: test
        path: test.parquet

lambda-chat

lambda-chat is a Japanese instruction-following dataset for supervised fine-tuning of chat models. It combines openly available datasets into one consistent chat format for easier use.

Purpose

The dataset is intended for training and evaluating Japanese chat and instruction-following models. Each example uses a list of messages with role and content fields.

Source Data

The dataset contains data from the following sources.

llm-jp/llm-jp-instructions and APTO-001/ja-safety-sft-dataset are published under CC BY 4.0 license.

Data Processing

The source data was deduplicated, normalized with NFKC, and filtered to remove code and LaTeX content. The examples were then randomly shuffled and split into train, validation, and test sets.

Dataset Size

Split	Rows
train	183,072
validation	1,868
test	1,868
total	186,808

The dataset contains about 50M assistant tokens measured by Gemma 4 12B tokenizer.

Columns

Column	Description
`id`	Unique identifier for the example.
`messages`	Chat messages. Each message has `role` and `content`.
`output_char`	Number of characters in all assistant responses.
`output_token`	Number of tokens in all assistant responses.
`dataset_name`	Source dataset name

Example

from datasets import load_dataset

dataset = load_dataset("KeisukeMiyamoto/lambda-chat")
print(dataset["train"][0])