Dataset Card for telecomadm1145/esjzone_2024

Dataset Details

Dataset Description

Novels from Esjzone.

Check telecomadm1145/creative_writing for instruction finetuning and telecomadm1145/esjzone_2024_chunked_8k for continuation.

• Curated by: telecomadm1145
• Language(s) (NLP): Chinese
• License: MIT

Dataset Sources

• Source Data: Publicly available online novels (especially light novels).

---

Uses

Better using DPP similar to below code:

from datasets import load_dataset
import random

hf_dataset = load_dataset("telecomadm1145/esjzone_2024_chunked_8k", split="train")

def make_prompt_response(example):
    text = example["content"]

    text = text[:3000] # TODO: Change if needed!

    split_ratio = random.uniform(1/6, 1/2)
    split_point = int(len(text) * split_ratio)
    
    context = text[:split_point]
    continuation = text[split_point:]

    user_message = f"""请按照这段文字的大致逻辑和剧情，以及人物发展，续写。直接输出后续的可能文段。
```text
{context}
```"""
    
    prompt = f"<start_of_turn>user\n{user_message}<end_of_turn>\n<start_of_turn>model\n" # Chat template maybe different
    
    return {
        "prompts": prompt,
        "responses": continuation,
    }

hf_dataset = hf_dataset.map(make_prompt_response)

---

Dataset Structure

• Format: JSON
• Splits:
  • train
• Fields:
  • book_title: Title of the book.
  • chapter_title: Title of the chapter.
  • content: Content of the chapter.