metadata

dataset_info:
  features:
    - name: message_tree_id
      dtype: string
    - name: prompter_message_id
      dtype: string
    - name: assistant_message_id
      dtype: string
    - name: instruction
      dtype: string
    - name: input
      dtype: string
    - name: output
      dtype: string
    - name: assistant_rank
      dtype: float64
    - name: assistant_detoxify
      struct:
        - name: identity_attack
          dtype: float64
        - name: insult
          dtype: float64
        - name: obscene
          dtype: float64
        - name: severe_toxicity
          dtype: float64
        - name: sexual_explicit
          dtype: float64
        - name: threat
          dtype: float64
        - name: toxicity
          dtype: float64
    - name: prompter_detoxify
      struct:
        - name: identity_attack
          dtype: float64
        - name: insult
          dtype: float64
        - name: obscene
          dtype: float64
        - name: severe_toxicity
          dtype: float64
        - name: sexual_explicit
          dtype: float64
        - name: threat
          dtype: float64
        - name: toxicity
          dtype: float64
    - name: assistant_labels
      struct:
        - name: count
          sequence: int32
        - name: name
          sequence: string
        - name: value
          sequence: float64
    - name: prompter_labels
      struct:
        - name: count
          sequence: int32
        - name: name
          sequence: string
        - name: value
          sequence: float64
  splits:
    - name: train
      num_bytes: 24671269
      num_examples: 12659
  download_size: 8255918
  dataset_size: 24671269
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: apache-2.0
task_categories:
  - text-generation
language:
  - ja
tags:
  - self-rewarding
  - oasst1
size_categories:
  - 1K<n<10K

Dataset Card for Dataset Name

を元に、self-rewardingのEFT(Evaluation Fine-Tuning data)の元データを作成しました。この後に、学習させたいモデルを使ってLLM-as-a-Judgeを行います。
Self-rewardingの論文では最終的に train: 1,630 records, test: 531 records に絞り込んでいます。

Dataset Details

Dataset Description

Curated by: HachiML
Language(s) (NLP): Japanese
License: Apache-2.0

Filtering Rule

以下のルールで絞り込んでいます。

First Conversational Turn
Single Turn Conversation
3パターン以上の回答を持つ
同一のparent_idを持つ回答パターンでrankに被りがない