Fine-tune an open-source model on internal documents to build a specialized assistant
Try the live demo to see the training workflow before setting up your own instance
Download a fine-tuned model in GGUF format to run locally with Ollama
| wenxin0319/llm-forge | albertaworlds/japanese-text-cleaner | ayangabryl/ngx-digit-flow | |
|---|---|---|---|
| Stars | 30 | 30 | 30 |
| Language | TypeScript | TypeScript | TypeScript |
| Setup difficulty | moderate | easy | easy |
| Complexity | 3/5 | 2/5 | 2/5 |
| Audience | pm founder | developer | developer |
Figures from each repo's GitHub metadata at analysis time.
Full stack setup needs Docker for Postgres plus separate frontend and backend installs, the hosted demo needs no setup.
LLM Forge is a web-based platform that lets you take an existing open-source AI model and train it on your own data, without setting up any servers or writing any training code. The idea is that you pick a model from a catalog, upload a file of examples, configure a few settings, and get back a custom AI model tuned to your specific topic or domain. The live demo is hosted at llm-forge-azure.vercel.app and works without a backend, so you can explore the interface before committing to anything. The platform is organized as a tabbed interface. The Model Catalog offers 14 pre-selected open-source models ranging from small ones that run on a laptop to larger ones suited for harder tasks. Each entry shows the memory requirements, what the model does well, and its license. Once you pick a model, you move to the Datasets tab to upload your training file. Supported formats include JSONL for instruction-answer pairs, CSV for row-based data, plain text documents, and Parquet files from data pipelines. Sample files covering medical records and clinical notes are included in the repository to let you test the flow immediately. The fine-tuning step is a three-step wizard. You choose your uploaded dataset, configure training settings like the number of epochs and learning rate (with sensible defaults already filled in), and then launch. The recommended training method is QLoRA, which the README describes as using 75 percent less GPU memory than standard approaches with minimal quality loss. A cost estimate is shown before you commit. While training runs, a monitor page displays a live loss curve, step-by-step logs, GPU utilization, and epoch progress. Training typically finishes in 10 to 60 minutes depending on data size and model choice. Finished models can be downloaded in several formats: raw adapter weights, a full merged model, a compressed GGUF file for local tools like Ollama, and a GPU-optimized GPTQ version. A one-click quantization button converts the result to GGUF format. The technical stack is Next.js on the frontend, NestJS with PostgreSQL on the backend, and JWT-based authentication. Both the frontend and backend deploy automatically on Vercel and Railway respectively. The code is MIT-licensed.
A web platform for fine-tuning open-source AI models on your own data without setting up servers or writing training code.
Mainly TypeScript. The stack also includes Next.js, NestJS, PostgreSQL.
Use freely for any purpose, including commercial use, as long as you keep the copyright notice.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly pm founder.
This repo across BitVibe Labs
Verify against the repo before relying on details.