README.md

metadata

license: mit
task_categories:
  - text2text-generation
language:
  - en
  - zh
configs:
  - config_name: chatgpt-2000
    default: true
    data_files:
      - split: train
        path: chatgpt-train-2000.jsonl
      - split: test
        path: chatgpt-test.jsonl
  - config_name: chatgpt-8192
    data_files:
      - split: train
        path: chatgpt-train-8192.jsonl
      - split: test
        path: chatgpt-test-8192.jsonl

Introduction

This repository holds the data file for translating TechLinked, which talks about mostly technology and science news.

Raw data is in the data/ folder. Scripts generate OpenAI's ChatCompletion Fine-tuning API formatted training data in jsonl format.

-2000 variants are designed to be used with GPT-3 with 8192 tokens context length limit. -8192 variants are designed to be used with GPT-4o mini with 128000 context window and 16384 max output tokens.

How to add data to this repository

0. pip install ass
1. Convert ASS file into .en.txt and .cn.txt files: python ./ass_extract.py [ASS Filename]
   This step will generate two files: Extracted - [Filename].en.txt and Extracted - [Filename].cn.txt
2. Move them into the data/ folder. You may want to rename them also, but keep their filenames the same except .en and .cn.
3. Run script: python ./generate_chatgpt_varlen data --maxlen MAXLEN --test-ratio TEST_RATIO
   • data is the data directory.
   • MAXLEN is recommended to be a quarter of the context window, or a little bit less than maximum output tokens, whichever is smaller.
   • TEST_RATIO is the ratio of data to be reserved for testing. A decimal number. Default is 0.2 This will generate three files:
   • combined-{MAXLEN}.jsonl: Test+Train data.
   • chatgpt-train-{MAXLEN}.jsonl: Train data.
   • chatgpt-test-{MAXLEN}.jsonl: Test data.

The other scripts are deprecated.