Update README.md

README.md

@@ -1,3 +1,75 @@
----
-license: mit
----
+---
+license: mit
+task_categories:
+- text-generation
+language:
+- en
+tags:
+- agent
+size_categories:
+- 10K<n<100K
+---
+# LiteGPT Dataset
+
+This repository contains a synthetic conversational dataset designed for training lightweight GPT-style language models like LiteGPT. The dataset consists of user-assistant dialogues with enriched prompts and responses.
+
+---
+
+## Dataset Description
+
+- **Format**: Each line in `corpus.txt` represents a single conversation in the following format:
+
+<BOS> <user>: <user_input> <assistant>: <assistant_response> <EOS> ```
+Special Tokens:
+
+<BOS>: Beginning of sequence
+
+<EOS>: End of sequence
+
+<user>:: Marks the user's input
+
+<assistant>:: Marks the assistant's response
+
+<PAD>: Padding token for fixed-length sequences
+
+Number of Conversations: At least 25,000 generated examples
+
+Content: The conversations include a variety of topics such as greetings, jokes, advice, AI knowledge, science questions, history, coding, and small talk.
+
+Dataset Generation
+The dataset is generated automatically using lite_gpt.py:
+
+python
+Copy code
+from lite_gpt import create_synthetic_corpus
+
+create_synthetic_corpus()
+This will:
+
+Randomly select a user prompt from a predefined list.
+
+Randomly select a corresponding assistant reply from a predefined list.
+
+Save the generated conversations into data_v2/corpus.txt.
+
+Directory Structure
+bash
+Copy code
+data_v2/
+└── corpus.txt     # Synthetic conversational dataset
+Tokenization
+The dataset is designed for GPT-2 tokenization.
+
+Each conversation is tokenized and padded to a maximum length (MAX_LENGTH) for model training.
+
+Special tokens are added to distinguish user and assistant turns.
+
+Usage
+Can be used to train lightweight language models.
+
+Supports sequence chunking for longer conversations.
+
+Works with any PyTorch-based GPT-style model.
+
+License
+This dataset is generated synthetically and is free to use under the MIT License.