FAQ¶
Short answers to the questions people ask most, with links to the details.
Cost and accounts¶
Is Tuneline Studio free?¶
Tuneline Studio is open source (Apache 2.0), and it does not bill you for compute. You pay the providers you choose, directly, for what you use. You can also train for free: on Kaggle's free GPUs, in a free Colab notebook, or on your own computer. See Hardware and costs.
Who pays for compute?¶
You do, through your own accounts. Runs on Modal, RunPod or Hugging Face Jobs, endpoints, and Hub models in the Playground are billed by that provider to the account whose key you saved. Tuneline Studio never pays for or resells GPU time. Set Stop at cost (USD) in the Run dialog to cap a run. See Running pipelines.
Which keys do I need?¶
A Hugging Face token with the Write role for almost everything, plus a key for where you run: Kaggle, Modal or RunPod (none for Notebook or My computer). See API keys and security.
Why is a RunPod key requested when I run on Kaggle?¶
Because the pipeline deploys an endpoint on RunPod at the end. The Sanctions alert triage template ends with a Deploy endpoint block set to RunPod Serverless. If you don't want an endpoint yet, select that block and press Delete node, or pick another provider in Where to serve. See Troubleshooting.
Hardware¶
Can I train without a GPU?¶
Only very small models. The CPU smoke test template trains a 135M-parameter model for 20 steps on any computer in a few minutes, to check your setup. Anything useful needs a GPU; free ones are available on Kaggle (two T4s, about 30 GPU hours a week, phone-verified account) and in Colab through the Notebook option.
What GPU do I need?¶
It depends on the model size, the method and the sequence length. As a rough guide, QLoRA fits 7B-class models on a single 16 to 24 GB GPU, LoRA needs about 2 to 3 times more memory, and full fine-tuning needs about 16 to 18 bytes per parameter, so only small models. The Run dialog and the Hardware page show, for your exact pipeline, which GPUs fit and the estimated time and cost. See Hardware and costs.
Does it work on a Mac?¶
The web app works in any modern browser. To train on a Mac, connect it as My computer: the macOS installer sets up PyTorch for Apple silicon, which uses the Apple GPU. QLoRA (bitsandbytes) is not available on macOS, so training uses LoRA. For larger models, use a cloud GPU. See Your own computer.
Can I use my own NVIDIA GPU?¶
Yes. Connect the computer once in Settings → Your computers, then pick My computer in the Run dialog; runs start within 30 seconds. On Windows, QLoRA is most reliable under WSL2. See Your own computer.
Data and models¶
Can I use my own dataset?¶
Yes. Upload a CSV, JSONL, JSON, Parquet or ZIP (images) file on the Datasets page, or use a Hugging Face dataset or a URL. Tuneline Studio detects the format (chat messages, ShareGPT, Alpaca, prompt/completion, preference pairs, plain text, vision) and shows a preview. See Datasets.
Which models are supported?¶
The Models page has a catalog of 237 open models that can be fine-tuned (Qwen, Llama, Gemma, Mistral, Phi, DeepSeek, Granite, OLMo, SmolLM, InternVL and more) with their license, gating and memory needs. You can also search the Hugging Face Hub or paste any repository id that Transformers can load, text or vision. See Models.
What is the difference between LoRA, QLoRA and full fine-tuning?¶
| Method | What is trained | Memory | When to use |
|---|---|---|---|
| QLoRA | A small adapter on top of a 4-bit copy of the model. | Lowest. | The default; 7B+ models on one GPU. |
| LoRA | A small adapter on a 16-bit copy of the model. | About 2 to 3 times QLoRA. | Slightly better quality when memory allows. |
| Full fine-tune | Every weight. | Highest (about 16 to 18 bytes per parameter). | Small models only. |
LoRA and QLoRA produce an adapter that needs its base model; Merge adapter turns it into a standalone model. See the block reference.
SFT, DPO, ORPO, SimPO, KTO or GRPO: which one?¶
| Block | Method | Data it needs | Use it to |
|---|---|---|---|
| Fine-tune (SFT) | Supervised fine-tuning | Example conversations or prompt/answer pairs. | Teach a task and an answer format. Start here. |
| Preference tuning (DPO / ORPO / KTO) | DPO, ORPO (no reference model), SimPO (reference-free) | A prompt with a chosen and a rejected answer. | Prefer better answers over worse ones, usually after SFT. |
| KTO | A prompt, an answer and a good/bad label. | The same, when you only have thumbs up or down, not pairs. | |
| Reinforcement tuning (GRPO) | GRPO with rule-based rewards | Prompts with a reference answer that can be checked. | Tasks with checkable answers, such as math, extraction or strict formats. |
| Vision fine-tune (SFT) | SFT for vision-language models | Images with questions and answers. | Documents, receipts, IDs, screenshots. |
Where does my model go?¶
To your own Hugging Face account, as a private repository by default. Training, merge and GGUF export blocks push there when their push setting is on (for example Push the adapter to Hugging Face, on by default), and Publish and register always does. The Registry keeps track of versions; the files stay on Hugging Face.
Is my data private?¶
Your pipelines, runs, datasets and keys are visible only to your account. Uploaded files are in a private storage bucket, keys are encrypted in Supabase Vault, and models go to private repositories by default. The providers you choose to run on process your data on their machines, so pick ones whose terms fit your data. See API keys and security.
Can I share a model or a pipeline?¶
Not between accounts inside the studio: each account only sees its own work. To share a pipeline, download it from the YAML tab and send the file; the other person applies it in their own YAML tab (see Pipeline YAML). To share a model, give people access to its Hugging Face repository, or make it public on Hugging Face.
Can I keep training a model I already trained?¶
Yes. Add a Registered model block and pick the version from your Registry, or in Base model set Start from LoRA adapter (optional) to your adapter's repository. Then connect a training block as usual.
Runs¶
Can I stop a run?¶
Yes. Press Cancel on the run page. It stops at the next safe point (training checks after every step), and steps that already finished stay cached for next time. See Watching runs.
Do I have to keep the browser open?¶
No. Cloud runs and runs on your computer continue when you close the tab, and the run page catches up when you open it again. A Notebook run lasts as long as the notebook keeps running, and a My computer run as long as the worker keeps running.
What does "cached" mean?¶
A step is Cached when its settings and inputs are exactly the same as in an earlier run, so Tuneline Studio reuses that run's outputs instead of computing them again. Change a setting or the data and the step (and everything after it) runs again. Turn it off with Reuse cached steps in the Run dialog. See Core concepts.
How long does a run take?¶
The CPU smoke test takes a few minutes. Real fine-tuning runs take from minutes to hours, depending on the model, the data size, the number of epochs and the GPU. The Run dialog estimates time and cost for each GPU before you start.
Why did my run fail at the quality gate?¶
The model scored below one of the gate's rules, so nothing after the gate ran. The error names the rule and the score, for example "Quality gate failed: label_recall_escalate = 0.95 (needs >= 0.98)". See Evaluation and quality gates.
Using your model¶
How do I use my trained model?¶
Three ways, from free to managed:
- On your computer with Python: load the adapter from your Hugging Face repository on top of its base model with Transformers (the model card has a How to use section). The repository's example app
examples/alert-triage/app.pydoes this for the triage template. - In Ollama, LM Studio or llama.cpp: add Merge adapter, Export GGUF and Ollama package to the pipeline, as in the Chat assistant (SFT → GGUF → Ollama) template.
- As an API: deploy it as an OpenAI-compatible endpoint on RunPod, Hugging Face or Modal (paid). Then call it from any OpenAI client, or chat with it in the Playground.
See Deployments.
Can I compare my model with the original?¶
Yes. Evaluate on test set scores the base model too (Also score the base model, on by default) and shows the change for every metric. In the Playground, Side by side lets you chat with two models at once. See Evaluation and quality gates.
Can I log to Weights & Biases?¶
Yes. Save your W&B key in Settings, then set Also log to to Weights & Biases on the training block. Charts still appear on the run's Metrics tab.