tiny-llm-dataset / README.md
Neural-Labs's picture
Add dataset
3204b96 verified
|
Raw History Blame Contribute Delete
1.24 kB
metadata
license: mit
language: en
task_categories:
  - text-generation
tags:
  - synthetic
  - conversational
  - greetings
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.txt

Tiny LLM greetings dataset

The training text for Neural-Labs/tiny-llm, an 824,120-parameter character-level language model by Neural Labs.

  • 12,000 short exchanges, about 697,000 characters, 56 distinct characters
  • 19 greeting topics: hello, how are you, what is your name, who made you, thanks, goodbye and so on
  • fully synthetic: every line is produced by make_data.py from hand-written phrase lists, with small random changes to capitals, punctuation and "i am / i'm"
  • no scraped, copied or personal data

Format

Plain text. Each exchange is two lines, and exchanges are separated by a blank line:

User: good morning
Bot: Good morning! Have a lovely day.

User: who made you
Bot: Neural Labs made me. They trained me on a small list of greetings.

Files

File What it is
train.txt The training text
make_data.py The script that generates it (python make_data.py writes the same text as input.txt)