Frequently Asked Questions

In Early Access: €4.99/month or €39.99/year (equivalent to €3.33/month). The price increases to €9.99/month after the first 100 users. Those who join now keep the lower price permanently.

All major credit and debit cards via Stripe: Visa, Mastercard, American Express, Discover. Additionally SEPA direct debit for Europe, Apple Pay, and Google Pay.

Yes, at any time, effective at the end of the current billing period. No notice periods, no hidden fees. Access remains active until the end of the paid period.

There is no blanket money-back guarantee: since this is digital content made available immediately, the statutory right of withdrawal is limited. If you want to try before you buy, use a free-months code at /redeem – "Redeem without subscription" needs no payment method and ends automatically. Our support team will help with any issues.

The price increases to €9.99/month for new users. Existing users keep their Early Access price permanently – even as the general price rises.

Not yet, but planned. Get in touch by email – we try to find solutions on a case-by-case basis.

Yes. During checkout click "Have a promo code?" or open the redeem page directly. Enter your code, we validate it and unlock your benefit instantly. You get codes through promotions, partners, or directly from us.

Three: discount codes (e.g. 50% off), free-months codes (e.g. 2 months free), and lifetime codes (permanent free access without any subscription).

You have two options. "Redeem without subscription" gives you full access for the free period with no payment method – access then ends automatically. "Start with subscription" stores your card with Stripe but only charges it after the free period, so it continues seamlessly. Cancel before then and you pay nothing.

A lifetime code unlocks your account permanently – no subscription, no payment, no expiry date. All updates included.

If you redeemed "without subscription", access ends automatically on the shown date and your API key is deactivated. Your dashboard shows you when that is in good time – there you can subscribe or redeem another code anytime to keep using FrameTrain.

Yes. If you subscribe during active free months, we carry over the remaining free time – the first charge only happens after it ends. Nothing is charged twice and no free time is lost.

Discount codes work with both the monthly and yearly plan; the discount is applied automatically at the Stripe checkout. Time-limited discounts (e.g. "for the first months") fit the monthly plan best – on the yearly plan the discount only applies to the first yearly invoice.

Yes, CPU training is possible but extremely slow. For productive work we recommend at least an NVIDIA GTX 1060 (6 GB VRAM) or Apple M1.

Yes. FrameTrain natively supports Apple M1, M2, M3, and M4 via Metal Performance Shaders (MPS). Training runs on the M-chip's GPU without requiring CUDA.

All NVIDIA GPUs with CUDA support from the GTX 10 series (Pascal architecture) onward. Tested: GTX 1060, 1080, RTX 2080, 3060, 3090, 4070, 4080, 4090. Recommendation for 7B models: RTX 3060 (12 GB) or better.

Multi-GPU training is on the roadmap but not yet implemented. FrameTrain currently supports one GPU per training session.

CUDA is NVIDIA's proprietary compute framework, very mature and fast. MPS (Metal Performance Shaders) is Apple's equivalent for M-chip Macs, slightly slower than CUDA but perfectly adequate for most fine-tuning tasks.

LoRA (Low-Rank Adaptation) is a method for fine-tuning language models with little VRAM. Instead of updating all model parameters, LoRA only trains small additional matrices. Recommended for: GPU with <24 GB VRAM, domain-specific adaptation, multiple adapters on one base model.

QLoRA combines 4-bit quantization of the base model with LoRA training. This reduces VRAM requirements by about 40–60%. QLoRA is the better choice when you have limited VRAM (under 12 GB) or want to fine-tune a 13B+ model.

Yes. FrameTrain supports CSV, JSON, and JSONL as dataset formats. You can also load datasets directly from the HuggingFace Datasets Hub.

Full fine-tuning updates all model parameters – needs lots of VRAM (40–80+ GB for 7B) but is more flexible. LoRA only trains small adapter matrices, needs much less VRAM (6–18 GB for 7B), and delivers equivalent results for most use cases.

Yes. Since version 1.3.4 there is the Dataset Builder with seven project types: mark objects (boxes, YOLO export), class per image, per text, per recording and per video segment, text → answer, and transcripts. You collect data from folders, the microphone, the web or let the AI assistant generate examples, label with the keyboard and export ready for training – split into train, val and test.

Yes, in the Dataset Builder via single URLs, whole websites (with depth and page limit) or sitemaps. robots.txt is respected, every server gets at least one second between requests, and every file keeps its URL, source page and stated license (PROVENANCE.csv). Whether you may use the content for training is up to you to check against the license. Pages that load their content via JavaScript yield little so far.

No. A model you have already trained can suggest boxes, classes, transcripts (Whisper) or answers. The “Uncertain” filter shows the least confident suggestions first, and nothing counts as confirmed until you press Enter. On top of that, wrong and uncertain results from the Laboratory go back into a builder project as new work.

No. Your training data never leaves your machine. All training runs locally. Only at startup does the app check the API key against our servers – no data or model weights are transferred in that process. The only exception: if you use the optional AI features with Claude, OpenAI or Groq, the context of your request (e.g. configuration, figures or script code) goes to that provider. With Ollama, that stays local too.

Yes. Since no training data leaves your own infrastructure, FrameTrain is ideal for GDPR-compliant ML projects. There is no telemetry and no hidden data transfers. Payment data is processed exclusively by Stripe (PCI-DSS Level 1).

No. All trained models and checkpoints are stored exclusively locally on your device. FrameTrain has no cloud storage for models.

No automatic telemetry. FrameTrain does not collect usage data, training data, or model metrics. The only default network communication is license validation (API key check) at startup. You can optionally opt in to anonymized crash/error reporting in the app's settings to help us fix bugs – off by default, never includes training data, and can be disabled at any time. The AI features are off by default as well; if you enable them with a cloud provider, requests go directly to that provider, not to FrameTrain.

Payment data is processed exclusively through Stripe, a PCI-DSS Level 1 certified payment service provider. We do not store any credit card data ourselves.

All PyTorch-based HuggingFace models: LLMs (Llama, Mistral, Phi, GPT-2), text classifiers (BERT, RoBERTa, DistilBERT), sequence-to-sequence models (T5, BART), and many more.

PyTorch models (.pt, .bin), SafeTensors (.safetensors), and GGUF format (.gguf). Import from the HuggingFace Hub works directly via model ID.

In the Model Manager, simply enter the HuggingFace model ID (e.g. meta-llama/Llama-3-8B) and click "Download". FrameTrain downloads all necessary files and sets the model up automatically.

Yes. In the Model Manager you can also provide the path to a locally stored model. Supported formats: HuggingFace format (model.safetensors / config.json), GGUF.

In version management you export any model version to your Downloads folder with one click – as a copy of all the files the training run produced, for HuggingFace models e.g. SafeTensors weights, config.json and the tokenizer. There is currently no built-in GGUF export. Everything stays on your machine; nothing goes to the cloud.

Training loss measures the model's error on the training data; validation loss measures it on a separate test set. If the validation loss is much higher than the training loss, you have overfitting.

Live metrics: training loss, validation loss, accuracy, learning rate, GPU utilization (VRAM, temp), elapsed/estimated time, gradient norm. All metrics are displayed as interactive charts.

Gradient clipping limits the maximum magnitude of gradients to prevent exploding gradients. FrameTrain sets max_grad_norm=1.0 by default. Reduce this value when you see loss spikes.

Yes. All training sessions are saved locally in the training history. You can retrieve and compare metrics from past training runs.

1. Download the .dmg file. 2. Open the DMG and drag the app to Applications. 3. Remove the quarantine flag: sudo xattr -cr '/Applications/FrameTrain.app'. 4. Launch the app. Detailed guide: /install

1. Download the .msi or .exe file. 2. Right-click → Properties → check "Unblock". 3. Run the installer. 4. If SmartScreen warns you: "More info" → "Run anyway". Detailed guide: /install

1. Install libfuse2: sudo apt install libfuse2t64 (Ubuntu 22.04+). 2. Download the .AppImage file. 3. Make it executable: chmod +x FrameTrain.AppImage. 4. Run: ./FrameTrain.AppImage. Detailed guide: /install

The API key is your license key for the desktop app. You receive it after purchase on the payment page. It is entered once when the app starts and then stored locally.

Common causes: 1. API key not entered or invalid. 2. No internet connection for license validation. 3. macOS: quarantine flag still active (sudo xattr -cr). 4. Windows: file not unblocked (Properties → Unblock). For further help: framesphere@gmx.net

FrameTrain checks for a new version at startup. On Windows and Linux the app downloads the signed update on request, installs it and restarts. On macOS it tells you about the new version and you get the download from the dashboard. You can check manually at any time under Settings → Updates.