Train AI on Your Documents.
Entirely on Your Machine.

FORGE creates custom AI models specialized for your domain — your terminology, your document types, your workflows. No cloud uploads. No machine learning skills needed.

FORGE main screen — hardware card & dataset view

Generic AI Doesn't Know Your Documents

Generic AI models are trained on the internet. They know a little about everything, but they don't know your documents.

When an insurance professional asks about "Coverage A: Dwelling," a generic model treats it as plain text. It doesn't know that this is the primary coverage section in a homeowner's policy, that sub-limits modify the main limit, or that exclusions and endorsements interact with it in specific ways.

The same is true for legal contracts, financial statements, engineering documentation, and clinical protocols. Your domain has structure, terminology, and interpretation patterns that generic models miss.

FORGE solves this by turning your documents into training data and building a custom model that understands your specific domain.

Six Steps to a Smarter Model

From document folder to domain-tuned AI — no machine learning knowledge required.

1

Select your documents

Pick any folder in your Librarian library as the training source.

2

Generate training data

FORGE analyzes your documents and creates question-answer pairs automatically.

3

Review and refine

Accept the good pairs, reject the weak ones. You control what the model learns.

FORGE model pairs Pair review panel screenshot
4

Choose a base model

Pick from our recommended models — or bring your own from HuggingFace.

5

Train

Click Start. Watch real-time progress. Training runs on your GPU — nothing leaves your machine.

FORGE model train Training progress screenshot
6

Assign to a folder

Attach your trained model to any document folder. It activates automatically when you chat.

Domain expert llm Folder assignment screenshot

Why FORGE Changes the Game

Domain Intelligence

Your model doesn't just search your documents — it understands them. It learns the terminology, structure, and interpretation patterns specific to your field.

Completely Private

Training runs locally on your hardware. Your documents, your training data, and your custom model never leave your machine. No cloud. No third parties.

No ML Skills Needed

FORGE handles the machine learning. You review question-answer pairs and click Start. Default settings are optimized for your hardware. Advanced users can tune parameters if they choose.

Gets Better Over Time

As you add more documents and retrain, your model improves. The investment compounds — your AI gets smarter the longer you use it.

What You Need

FORGE requires an NVIDIA GPU for training. Before you start, the app detects your hardware and tells you exactly what's possible.

Your GPU VRAM What You Can Train Example Cards
Entry-level 8 GB Models up to 7B parameters GTX 1070, RTX 3060, RTX 4060
Mid-range 12–16 GB 7B models with faster training RTX 3060 Ti, RTX 4070, RTX 4080
High-end 24+ GB Larger models with full performance RTX 3090, RTX 4090, Tesla/A-series

Don't know your GPU specs? FORGE automatically detects your hardware when you open it — no guesswork needed. It shows your GPU model, available VRAM, and which base models will fit.

NVIDIA GPUs only. FORGE training requires an NVIDIA GPU with CUDA support. This covers the GTX 1070 and newer, all RTX cards, and Tesla/A-series datacenter GPUs. AMD and Apple Silicon support is on our roadmap.

Under the Hood

The technical details for ML practitioners and power users.

Training Technique

LoRA (Low-Rank Adaptation) with QLoRA 4-bit quantization via bitsandbytes. Efficient fine-tuning that runs on consumer GPUs.

Output Format

GGUF — optimized for local inference. Models are registered with Ollama and appear alongside your base models.

Base Models

Curated recommended list, plus any HuggingFace-compatible model via advanced mode.

Configurable Parameters

Rank, alpha, learning rate, and epochs are all adjustable. Sensible defaults are provided and optimized for your detected hardware.

CUDA Stack

torch==2.6.0+cu124 built for CUDA 12.4. Hardware detection via torch.cuda.is_available(). NVIDIA GPUs with 8 GB+ VRAM.

Crash Isolation

Training runs in a separate process. If training fails, Librarian continues unaffected. Reviewed training data is preserved across restarts.

FORGE Add-On

Available with any Personal or Team subscription.

Add-On
€99 /month

FORGE is a monthly add-on to your existing Librarian subscription. Train unlimited models on your own data using your hardware.

View Plans & Pricing

Requires an NVIDIA GPU with at least 8 GB VRAM.

Common Questions About FORGE

Do I need machine learning experience?

No. FORGE handles the entire technical process — from generating training data to configuring model parameters. You review question-answer pairs and click Start. That's it.

Does my data leave my machine during training?

No. All training runs locally on your GPU. Your documents, training data, and the resulting custom model never leave your machine. No cloud processing, no uploads, no third parties.

How long does training take?

It depends on your hardware and dataset size. Typically between 1–4 hours. FORGE shows a time estimate before you start so there are no surprises.

Can I use my trained model outside Librarian?

Yes. Models are exported in GGUF format and registered with Ollama, so any tool that reads Ollama models can use them.

What if training fails or crashes?

Training runs in a separate process — your active Librarian sessions are never interrupted. You can restart training without losing your reviewed data.

What models can I use as a base?

FORGE provides a curated list of recommended models optimized for different use cases. Advanced users can also use any compatible model from HuggingFace.

Does FORGE work on Mac or AMD GPUs?

Not currently. FORGE requires an NVIDIA GPU with CUDA support. The entire training stack — including PyTorch CUDA 12.4, QLoRA quantization, and hardware detection — depends on NVIDIA's CUDA toolkit. AMD (ROCm) and Apple Silicon support is on our roadmap.

How many models can I train?

There's no limit. Train as many models as you like and assign each one to different document folders for domain-specific intelligence.

Ready to Train Your Own Model?

Start with a free trial of Librarian, then add FORGE to build AI that truly understands your documents.

Start Free Trial Product Overview