Instructions to use MainStack/marvy-1-14B-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use MainStack/marvy-1-14B-lora with PEFT:
Task type is invalid.
- MLX
How to use MainStack/marvy-1-14B-lora with MLX:
# Make sure mlx-lm is installed # pip install --upgrade mlx-lm # if on a CUDA device, also pip install mlx[cuda] # Generate text with mlx-lm from mlx_lm import load, generate model, tokenizer = load("MainStack/marvy-1-14B-lora") prompt = "Once upon a time in" text = generate(model, tokenizer, prompt=prompt, verbose=True) - Notebooks
- Google Colab
- Kaggle
- Local Apps
- LM Studio
- MLX LM
How to use MainStack/marvy-1-14B-lora with MLX LM:
Generate or start a chat session
# Install MLX LM uv tool install mlx-lm # Generate some text mlx_lm.generate --model "MainStack/marvy-1-14B-lora" --prompt "Once upon a time"
Licensing — marvy-1-14B
marvy-1-14B uses a layered (dual) license that reflects what is built on top of an upstream open model versus what MainStack authored.
| Component | License | What it covers |
|---|---|---|
Model weights (*.safetensors, GGUF quants, LoRA adapter) |
Apache-2.0 | The fine-tuned weights. These are a derivative of Qwen2.5-14B-Instruct (Apache-2.0); per that license they remain Apache-2.0 and free to use, modify, and redistribute. |
| MainStack original contributions | CC-BY-4.0 | The model cards, documentation (USAGE.md, VALIDATION.md, benchmark), the benchmark charts, the curated training-data methodology, and the pipeline framing authored by MainStack. |
What this means in practice
You may (under Apache-2.0, for the weights)
- Use marvy-1-14B commercially, privately, or in research.
- Fine-tune, distill, quantize, merge, or otherwise build on the weights.
- Redistribute the weights, including modified versions.
…provided you retain the NOTICE file in derivatives and redistributions
(Apache-2.0 §4(d) — this is mandatory and carries the attribution request).
You must (under CC-BY-4.0, for our contributions)
If you reuse MainStack's documentation, model cards, benchmark, or charts — e.g. copying our eval methodology, reproducing our charts, or lifting card text into your own model — you must give attribution: credit "MainStack" and link to https://huggingface.co/MainStack/marvy-1-14B. This is a binding condition of CC-BY-4.0, not just a request.
We ask (attribution for the model)
If you use marvy-1-14B as a baseline, a starting point for your own fine-tune,
a distillation source, or an evaluation comparison, please credit MainStack
and cite the entry in the model card. See NOTICE and the card's Citation
section.
Why the weights can't be more restricted
Qwen2.5-14B-Instruct is released under Apache-2.0, which grants every recipient an irrevocable, royalty-free right to use and redistribute. A fine-tune cannot revoke those rights on the resulting weights. MainStack's protection therefore lives where it legally can: (1) the CC-BY-4.0 license on our own authored materials, and (2) the Apache-2.0 NOTICE that must travel with the weights.
Files
LICENSE— Apache-2.0 (governs the weights; inherited from the base model).LICENSE-CC-BY-4.0— CC-BY-4.0 (governs MainStack's documentation and other original contributions).NOTICE— required attribution notices (retain in derivatives).CITATION.cff— citation metadata.