md2LLM
A local-first management platform that turns personal Markdown archives into training data, hardware-aware fine-tuning workflows, and chat-ready open-source models.
md2LLM turns personal Markdown notes into a local-first fine-tuning workflow that covers dataset generation, base-model selection, hardware-aware training decisions, and local chat after training instead of stopping at a single training script.
The product supports `4` dataset-generation goals and `3` training routes, which makes it practical for different model behaviors, machine limits, and starting points across the workflow.
Inspiration Behind It
The project comes from the reality that years of project notes, research logs, and learning documents often stay trapped in Markdown folders instead of becoming part of an interactive system. md2LLM treats that personal writing as a trainable knowledge base rather than passive storage.
It was built around the pain point that fine-tuning becomes hard long before the actual training command. Users need help shaping datasets, deciding how the model should sound, understanding whether their machine can handle local training, and keeping records of each run so results can be compared over time.
System Architecture
Architecture Diagram
md2LLM
11 nodes
About The Flow / Architecture
The system is split into a React and Vite frontend, a FastAPI backend, a Markdown-processing pipeline, and separate training backends for MLX and Unsloth. That keeps ingestion, data generation, training orchestration, and local inference from collapsing into one opaque script.
The workflow moves through explicit stages: ingest notes or existing JSONL, generate training pairs, inspect hardware capability, recommend a base model and training route, then expose the resulting model locally through Ollama. Artifacts are stored across output and model directories so runs remain inspectable after training finishes.
md2LLM is a management and observability layer for personal-model fine-tuning. It ingests Markdown notes, generates training-ready JSONL data, recommends a training path based on available hardware, and helps users move from raw notes to a locally usable model.
The repository combines a browser UI, FastAPI control layer, Markdown parsing pipeline, MLX and Unsloth training scripts, a Colab fallback, and Ollama-based local chat. It stands out because it focuses on the operational friction around fine-tuning, not just the training step itself.