Iro96/terrylm_60k_generic

TerryLM Chat Dataset

A training data use for TerryLM - a ~10M parameters LLM

Dataset Description

TerryLLM 60K Generic Chat Dataset is a curated conversational dataset designed for supervised fine-tuning (SFT) of large language models on multi-turn dialogue understanding and generation.

The dataset contains approximately 60,000 chat-style conversations across a variety of lightweight and open-ended topics. Each sample is structured as a structured dialogue between a user and an assistant, preserving full conversational context through ordered message histories.

Example

{
  "id": "train-000000",
  "topic": "play",
  "messages": [
    {"role": "user", "content": "do you want to play a guessing game"},
    {"role": "assistant", "content": "yes. can the cup be part of it"},
    {"role": "user", "content": "sure. what role should it have"},
    {"role": "assistant", "content": "it can be the brave little hero. every game needs one."}
  ]
}

Fields

Field	Type	Description	Example
id	string	Unique identifier for each conversation sample	train-000000
topic	string	High-level label describing the conversation theme	play, memory, chat
messages	list of structs	Ordered multi-turn conversation between user and assistant	[{"role":"user","content":"..."}, {"role":"assistant","content":"..."}]

Generation

Data is synthetically generated using template composition with randomized components for high output diversity.

License

MIT