DadAI is an open-source project built to support new fathers during pregnancy and early parenthood.
The idea is simple: provide emotionally intelligent, practical guidance powered by LLMs — and built on Mistral 7B.

Most resources around parenting are either mother-centric or scattered across forums.
As a first-time dad, I realized how hard it can be to find support that's both practical and emotionally relevant — and I wanted to make things easier for other future dads.

DadAI aims to provide a clear, AI-driven interface that supports:

• Emotional support during pregnancy
• Concrete actions and reminders
• Guidance on sleep, communication, and partner well-being

• Mistral 7B (quantized with GGUF)
• LoRA fine-tuning with QLoRA
• Python (Transformers, PEFT, Datasets)
• Deployment via LocalAI
• Web interface (optional – future)

dadAI/
├── data/                       # Datasets (raw, cleaned, formatted)
│   ├── reddit_dataset.jsonl
│   ├── cleaned_dataset.jsonl
│   └── formatted_dataset.jsonl
├── lora_finetune/             # Fine-tuning and inference
│   ├── train.py
│   ├── merge_lora.py
│   ├── inference.py
│   ├── inference_batch.py
│   ├── prepare_dataset.py
│   └── convert_to_gguf.py
├── scripts/                   # Data collection, formatting, tests
│   ├── collect_reddit_data.py
│   ├── format_reddit_data.py
│   ├── clean_dataset.py
│   ├── check_dataset_format.py
│   ├── test_reddit_connection.py
│   └── show_random_sample.py
├── models/                    # LocalAI-compatible GGUF models
├── tests/                     # Prompt examples, screenshots
│   └── Prompt dadAI.png
├── requirements.txt
├── .env                       # PRAW credentials
├── .gitignore
└── README.md

Fine-tuning and inference will be tested on RunPod using QLoRA and Mistral 7B.

• Train an assistant on real parenting data (Reddit, BabyCenter)
• Optimize responses via QLoRA and test performance locally
• Package the assistant behind a simple OpenAI-compatible API

• Model: TheBloke/Mistral-7B-Instruct-v0.1-GPTQ

• Quantization: GPTQ 4-bit

• Fine-Tuning: QLoRA + PEFT (LoRA adapters)

• Data: 400+ high-quality Reddit pairs (Instruction/Response)

• Output: LoRA weights (~100MB) + merged model planned

• inference.py: basic text generation

• inference_batch.py: batched inputs

• dadAI_inference_test.ipynb: sandbox notebook

You can run DadAI locally using LocalAI, an open-source alternative to the OpenAI API.

This setup uses the Mistral 7B Instruct model in GGUF format and exposes a local /v1/chat/completions endpoint, fully compatible with OpenAI’s API.

We recommend downloading the Q4_K_M quantized version from TheBloke's Hugging Face page:

wget https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.1-GGUF/resolve/main/mistral-7b-instruct-v0.1.Q4_K_M.gguf -P localai/models/

⚠️ This file is not included in the repo (see localai/models/README.md for details)

More to come soon:

• Fine-tuning instructions (QLoRA)
• LangChain agent with memory
• Streamlit chatbot

Benoît Rossignol
📍 Geneva
💼 Solution Architect
🧠 AI & PreSales Leader