NELLA
A chat-driven LLMOps agent that turns a folder of documents into a fine-tuned, domain-specific language model, with human approval at each pipeline step.
NELLA automatically generates training data from uploaded documents, then selects a base model, trains it and evaluates the result, with the user approving each stage through a chat interface rather than configuring hyperparameters by hand. It runs locally on a personal computer or an institution's server rather than as a cloud service.
It ingests PDF, Word and Hangul (HWP) documents, automatically synthesizes question-answer and preference-tuning training data from them, and filters out low-quality examples with rule-based and LLM-based checks. Evaluation supports BLEU, ROUGE and LLM-judge metrics, with automatic data augmentation and retraining suggested from the results, and the finished model can be tested immediately in chat.
It also supports retrieval-augmented question answering over the uploaded documents as an alternative to fine-tuning. It ships a Windows installer that needs no separate Python, Node.js or Docker setup, as well as a Docker-based setup, and a hosted "agent" mode through a separate platform.