This document is not legal advice. Consult a qualified attorney for specific decisions.
Adapter files (.lora)¶
A .lora checkpoint produced by Juno contains delta weights derived from a base model and your
training data. Its legal status as a derivative work is unsettled and jurisdiction-dependent.
Conservative position: treat a .lora file as a derivative of the base model and apply the base
model’s license to its redistribution.
Merged GGUFs¶
./juno merge writes a new GGUF combining frozen base weights with adapter deltas. The
resulting file is more likely to be considered a derivative work of the base model than the
.lora adapter alone. Before redistributing a merged GGUF:
Confirm the base model license permits redistribution of derivative works.
Confirm your training data does not introduce additional copyright claims.
If the base model requires attribution, include it in any release artifact.
Models on which redistribution of merged outputs is known to be permitted under their standard license (as of 2026-06): Mistral 7B (Apache 2.0), Phi-3 (MIT).
Models requiring additional review before redistribution: LLaMA 3 (Meta license conditions), and any model with a non-commercial or prohibited-use clause.
Training data¶
Juno does not inspect training data. You are responsible for ensuring that data fed to the LoRA training pipeline does not infringe third-party copyrights and complies with the terms of any dataset license. Models trained on proprietary or licensed data may carry obligations that survive into the resulting adapter and merged weights.
See also¶
<- 9.2 Third-Party Model Weights | Table of Contents | 9.4 Contributor License Agreement ->