language:

• en
• ja
  tags:
• anime
• myanimelist
• conversational
• instruction-tuning
• fine-tuning
• recommendation
• uncensored
  size_categories:
• 1K<n<10K
  source_datasets:
• original
  license: mit

AniSage Anime Conversations Dataset

A dataset of 4,739 high-quality anime expert conversations generated from the
Anime Dataset 2026 — 5,000 Titles from MyAnimeList
Kaggle dataset using GPT-4o-mini.

What is this?

Each conversation is a question-answer pair where AniSage — a brutally honest,
opinionated anime expert gives unfiltered takes on anime series. Real opinions backed by real data.

What's covered

• 4,771 anime titles spanning 1970–2027
• 71 question templates across 9 categories:
  • Recommendations & watchability
  • Ratings & MAL score debates
  • Hot takes & unpopular opinions
  • Pacing, filler & adaptation quality
  • Studio & production analysis
  • Genre comparisons
  • Character & story analysis
  • Rewatchability & legacy
  • Romaji & native title questions

Data source

Raw anime metadata sourced from MyAnimeList via the Jikan v4 API (scraped November 2026).
Includes titles, MAL scores, genres, studios, synopses, member counts, and ratings.

Format

Each row contains:

• system — AniSage persona prompt
• user — question about a specific anime
• assistant — AniSage's opinionated answer
• messages — full ChatML format list (ready for direct fine-tuning)

{
  "system": "You are AniSage...",
  "user": "Is Attack on Titan overrated or underrated?",
  "assistant": "Let's get one thing straight..."
}

Intended use

• Fine-tuning LLMs to be anime domain experts
• Instruction tuning with opinionated conversational data
• Anime recommendation systems
• Research into domain-specific LLM personality

Model trained on this dataset

Coming soon — Llama 3 8B fine-tuned on this dataset + additional instruction data.

Created by

Made by Sevenzoro321 as part of a learning
project to understand LLM fine-tuning pipelines end to end.

Dataset generation: GPT-4o-mini
Source data: MyAnimeList via Kaggle