machinelearnear/multiturn_chat_milei_gpt

Milei-GPT Dataset

Che y si queremos hacer un LLM que hable de la misma forma que un famoso ... como hacemos? Este repo es una excusa para aprender a preparar un dataset para fine-tunear algún LLM, aprender como evaluarlo, como tokenizarlo, como extenderlo de formar sintética, y tantas otras cosas. Al final, si todo sale bien, vamos a tener un modelo que va a hablar como la persona que elegimos, y le podemos poner un RAG (retrieval augmented generation) encima para que nos traiga un contexto correcto y factual en las respuestas. Por ahora, la idea es hacerlo sobre Llama3-8B y usando APIs públicas para procesar la data, sobre mas de 300 horas de entrevistas.

Paso a paso, que vamos a hacer

• Encontrar todas las entrevistas en YT de algún famoso/a
• Transcribir las entrevistas
• Preparar un dataset (convertir a ChatML, tokenization, data sintética, etc.)
• Elegir un modelo base eg. Llama3-8B o Phi-3-mini-128k-instruct
• Fine-tuning del LLM, evaluación del modelo, y push to HF.
• Armar un RAG indexando las entrevistas y meterle este LLM

Mas info en: https://github.com/machinelearnear/milei-gpt

Dataset details

Dataset type: The Milei-GPT Dataset is a collection of transcriptions and speaker diarizations from various interviews of Javier Milei. It is intended to facilitate the fine-tuning of large language models (LLMs) to emulate the speaking style and content of Javier Milei.

• huggingface_dataset.parquet contains the processed dataset with detailed metadata and speaker-aware transcripts.

Dataset date: The dataset was collected and processed in May 2024.

Project or resources for more information: https://github.com/machinelearnear/milei-gpt

License: Apache-2.0

Intended use

Primary intended uses: The primary use of the Milei-GPT Dataset is to train, fine-tune, and evaluate language models that emulate the speaking style of Javier Milei. This dataset is particularly useful for developing conversational AI systems and chatbots that require a specific speaking style and tone.

Primary intended users: The primary intended users of this dataset are researchers and developers in the fields of natural language processing, machine learning, and artificial intelligence, particularly those focusing on conversational models and dialogue systems.

Dataset Construction

Features:

• video_id: Unique identifier for each video
• channel_id: The ID of the YouTube channel
• channel: The name of the YouTube channel
• uploader_url: URL of the uploader's YouTube profile
• url: URL of the YouTube video
• title: Title of the YouTube video
• duration: Duration of the video in seconds
• view_count: Number of views of the video
• candidate_name: Name of the candidate (Javier Milei)
• messages: List of transcribed and diarized messages from the video

Collection Process:

1. Data Collection: Interviews of Javier Milei were collected from YouTube.
2. Transcription and Diarization: Audio was transcribed using WhisperX and diarized using NVIDIA NeMo to distinguish between different speakers.
3. Dataset Preparation: The dataset was structured and formatted to include detailed metadata and speaker-aware transcripts, ready for use in training and fine-tuning LLMs.

Intended Workflow:

1. Finding Interviews: Use scripts to gather interview data from YouTube.
2. Transcribing and Diarizing: Transcribe and diarize the audio to identify and separate different voices.
3. Preparing the Dataset: Structure the data for model training, including tokenization and synthetic data generation.
4. Model Selection and Training: Choose a base model (e.g., Llama3-8B) and fine-tune it with the prepared dataset.
5. Evaluation and Deployment: Evaluate the model's performance and deploy it with a RAG system for enhanced response accuracy.

Citation

If you use this dataset in your research, please cite it as follows:

@dataset{milei_gpt_2024,
author = {machinelearnear},
title = {Milei-GPT Dataset},
year = {2024},
url = {https://huggingface.co/datasets/machinelearnear/milei-gpt},
}