FAQ

The most common questions we hear, grouped by what you are usually trying to decide. If your question is not here, check the Reference troubleshooting page or the Glossary.

Getting started

Do I need a GPU to use Ertas?

No. Ertas runs entirely in the browser. Training happens on managed GPUs (T4 or A10G) that Ertas provisions on demand. You only pay for the GPU minutes your run actually uses.

Do I need to know Python?

No. Ertas's canvas is visual end-to-end. If you want to script around the platform later, an API and CLI are on the roadmap, but neither is required to build, train, evaluate, or export a model today.

What is the smallest dataset I can fine-tune on?

Ertas does not enforce a row-count minimum, but for effectiveness we recommend starting with at least 20 rows so the model has enough signal to converge. You will rarely get useful behaviour below 500 well-curated examples for an instruction-tuning task; the best free-tier first run is 1,000 to 5,000 rows. See Dataset quality.

Can I fine-tune on multi-turn conversations?

Yes. Use the conversations format described in JSONL format. Ertas detects the chat template from the base model and applies it automatically.

Pricing and credits

How does billing work?

Ertas bills by GPU-minute. Every run shows an estimated cost in the Training Confirm dialog before you press play. You spend credits only while the GPU is attached and training. Queue time and provisioning time are free.

What happens if my run fails?

If your base model is from the verified catalog, credits for failed runs are refunded automatically. If you brought an unverified Hugging Face model and acknowledged the compatibility warning, you accepted the credit risk yourself. See Handling failures.

What is included in the free tier?

The Free plan refreshes 5 credits per day (with a 30-credit monthly ceiling), supports one project, and includes 250 MB of dataset storage and 5 GB of model artifacts storage in Hub. It uses the T4 GPU, which can train any base model under 5B total parameters except Gemma 4 E2B. Gemma 4 E2B has 2.3B effective compute parameters but 5.1B total once its Per-Layer Embedding lookup tables are counted, which pushes its training memory footprint past T4. Free-tier-compatible picks include Phi-3 mini (3.8B), Gemma 3 1B / 2B / 4B, SmolLM, TinyLlama, Llama 3.2 1B / 3B, Qwen 2.5 0.5B / 1.5B / 3B, and similar.

That is enough to complete the Quickstart and run a handful of small experiments per day. Models 5B total parameters and larger (including Gemma 4 E2B and E4B) require A10G, which is gated behind paid plans (Builder, Pro, Business). The pricing page has the current per-plan details.

Can I cap my monthly spend?

Not yet. Billing caps (soft and hard) are on the roadmap. Until they ship, the practical guardrail is to watch the live credit accrual in the Run panel and stick to predictable Max-Steps-based runs while iterating. If you anticipate a long sweep, run a small calibration first and multiply.

Models and licensing

What models can I fine-tune?

The catalog covers Llama 3 / 4, Mistral and Mixtral, Phi-3 / 4, Gemma 3 / 4, Qwen 2.5 / 3, SmolLM, TinyLlama, and other popular open-weights families. You can also bring any Unsloth-compatible model from a Hugging Face URL. See the models index.

Can I use the model I fine-tune commercially?

The base model's license controls what you can do with the resulting weights. Most catalog models (Llama, Mistral, Phi, Qwen) allow commercial use with conditions. A few are research-only. The model picker shows the license, and each model page spells out the terms. Read the license before you ship.

Does Ertas keep a copy of my fine-tuned model?

Your trained artifacts (LoRA adapter, GGUF) live in your project storage. You can download them at any time, and you can delete them at any time. Ertas does not use your private weights or your private datasets to train any other model. See the Privacy policy.

Datasets and privacy

Where is my dataset stored?

Datasets you upload live in your project storage, encrypted at rest. They are visible only to your account and any teammates you grant access to.

Does Ertas train on my data?

No. Your datasets are used only to train the model you ask Ertas to train. They are not pooled into any shared training set, not used for analytics, and not visible to other customers.

Can I delete a dataset?

Yes. Open the Data Craft tab, open the dataset's detail view, and choose Delete. Deletion is immediate and removes the file. The training that already happened is not affected, but rerunning a job from the Runs tab that referenced the deleted dataset will fail. Model artifacts produced by past runs (LoRA, GGUF) live in Hub and are not affected.

What if my dataset contains PII?

Ertas does not strip PII automatically. You are responsible for redacting personal data before upload. The Dataset quality page covers a few quick redaction patterns.

Training

How long does a typical run take?

Most first runs in the Quickstart finish in 20 to 40 minutes. Larger models, longer datasets, or higher step counts can push this to 2 to 6 hours. The Training Config picker estimates total steps and runtime before you press play.

Can I run multiple training jobs in parallel?

Yes. There is no hard concurrency limit. Each Action Module on the canvas runs as an independent job. The practical limit is your credit balance: starting a run puts its estimated cost on hold, and new runs draw from the non-held pool. If you do not have enough non-held credits for the next run's estimate, it is blocked at submission time. See Parallel runs.

What if I need to stop a run early?

Open the Run panel, expand the run, and click Cancel. You are billed only for the GPU minutes used before cancellation. Cancellation is not reversible, but you can rerun the same config at any time.

Why is my loss not going down?

Most often: learning rate too low, dataset too small, or formatting mismatched. See the Datasets troubleshooting page for a systematic checklist.

Deployment

What do I get when training finishes?

By default, a Q4_K_M quantised GGUF file plus the raw LoRA adapter as a downloadable archive. The GGUF is ready to load into Ollama, llama.cpp, LM Studio, or any compatible runner.

Can I run my model on a phone?

Yes. The GGUF format runs on iOS (via llama.cpp or MLC), Android (via llama.cpp or MLC), and on every desktop platform. See the Ship section for platform-specific guides.

Do I need to send user data back to Ertas at inference time?

No. Once you have downloaded the GGUF, inference is fully local to wherever you load the model. Ertas's servers are not in the loop.

What if I want to host the model on my own server instead of on-device?

The GGUF works in any llama.cpp-compatible server runtime, including llama.cpp's HTTP server, Ollama in server mode, and vLLM (with conversion). Ertas itself does not host inference.

Troubleshooting

My GGUF download is failing.

Check that the run finished with status Completed. Failed and cancelled runs do not produce GGUFs. If the run completed but the download stalls, retry from the Run panel. See Verifying exports.

The picker says my model is locked.

The model requires a GPU tier above what your plan allows. Click the model card and Ertas surfaces an upgrade prompt that takes you to the Billing settings. See Credits and usage.

Ollama refuses to load my GGUF.

Update Ollama to the latest version. Older Ollama builds do not recognise the most recent quantization formats. If the file still fails to load, try llama.cpp directly to confirm the file itself is valid.

What's next