README.md

metadata

license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - poetry
size_categories:
  - 10K<n<100K

PromptPoetry

中文 | English

Introduction

Large language models (LLMs) are widely used for poetry generation, but synthesizing training data for that task from existing models can lead to reduced diversity or introduce the biases already present in those models.

To address this, this dataset collects real English-language poems and pairs each one with a natural-language generation prompt, so as to better train and evaluate models on the ability to "generate a poem from a given instruction."

Every sample is a two-turn conversation: the user describes the desired poem in natural language, and the assistant replies with the corresponding original poem.

Dataset Composition and Structure

Files

The dataset contains two JSON Lines data files.

File	Entries
data/poems_queries/poems_queries_multilingual.jsonl	9,291
data/Poetry-Foundation-Poems/Poetry-Foundation-Poems.jsonl	13,745

23,036 samples in total.

Field Structure

Each line is a standalone JSON object (JSON Lines format):

{
  "conversations": [
    {
      "role": "user",
      "content": "Write a haiku or short poem about the theme of loneliness. Use the image of childhood as a source of warmth to contrast with the present feeling of isolation."
    },
    {
      "role": "assistant",
      "content": "06 Haiku Alone\n\nloneliness\nhis childhood\nwarms him up\n..."
    }
  ]
}

• conversations — the two dialogue turns, being the user instruction and the original poem returned by the assistant.
• language — the language tag of the prompt, given as an ISO 639-1 language code (e.g. ja, ru). English and Chinese prompts have no such field; a sample mixing several foreign languages carries a list.

Data Samples

English prompt

{
  "conversations": [
    {"role": "user", "content": "Write a poem titled \"Objects Used to Prop Open a Window\"."},
    {"role": "assistant", "content": "Objects Used to Prop Open a Window\n\nDog bone, stapler,\ncribbage board, garlic press\nbecause this window is loose—lacks\nsuction, lacks grip.\n..."}
  ]
}

Non-English prompt

{
  "conversations": [
    {"role": "user", "content": "Écris un court poème anglais, cynique, sur le fait de porter trop de croix, avec peu d'espoir, et des gens qui prient un dieu auquel ils ne croient pas."},
    {"role": "assistant", "content": "All Alone With Out Hope 0\n\nSO many crosses,\nso little hope, so many\npuppets, praying for\nsome thing they dont\neven believe in...."}
  ],
  "language": "fr"
}

Language Distribution

The prompt language composition of poems_queries_multilingual.jsonl:

Language	Code	Entries
English	(no field)	7,320
Chinese	(no field)	862
French	fr	100
German	de	88
Russian	ru	85
Japanese	ja	84
Polish	pl	82
Italian	it	81
Portuguese	pt	79
Arabic	ar	77
Thai	th	76
Vietnamese	vi	75
Spanish	es	69
Hindi	hi	69
Indonesian	id	67
Korean	ko	66
Others (Danish, Bulgarian, Bengali, Norwegian, Croatian, Romanian, Welsh, Somali, Greek)	da bg bn nb/no hr ro cy/so el	9
Mixed-language prompts	ja+de, es+cy	2

There are 16 single-language tags, plus 2 mixed-language entries. English and Chinese prompts carry no language field, so they are not counted among the tagged entries.

Poetry-Foundation-Poems.jsonl is almost entirely English; only 21 entries carry a non-English tag (fr 7, de 7, es 5, ca 1, es+tl 1).

Construction

Poetry-Foundation-Poems.jsonl

The source data comes from the PoetryFoundationData.csv provided by suayptalha/Poetry-Foundation-Poems (a scrape of the Poetry Foundation website). During conversion the prompt is assembled from a fixed template:

Write a poem titled "TITLE" about tag1, tag2, tag3.

The tags are the poem's first 3 tag fields; a poem without tags keeps the title only. The assistant reply is formatted as TITLE\n\nPOEM, taken verbatim from the CSV.

poems_queries_multilingual.jsonl

The original poems are first scraped from the dataset AJMC2002/poems, then the model minimax-m3 is called to generate diverse and multilingual user prompts, in order to strengthen the model's ability to follow multilingual instructions.

Cleaning

All the data has been carefully cleaned, including repairing mojibake, removing empty-body lines, etc.

Usage

The data is released in JSON Lines format and can be read directly:

import json

with open("data/poems_queries_multilingual.jsonl", encoding="utf-8") as f:
    for line in f:
        if line.strip():
            sample = json.loads(line)
            conversations = sample["conversations"]
            # conversations[0]["content"] is the prompt
            # conversations[1]["content"] is the original poem
            # sample.get("language") is the ISO 639-1 prompt tag, if any

Issues and Limitations

Please report any issues or suggestions.

MiYago Datasets

MiYago Datasets is a dataset collection built specifically to improve the creative writing and roleplay capabilities of models.

License

This dataset is licensed under the Apache License 2.0.