Frequently Asked Questions About Radal
What model architectures does Radal support?
Radal focuses on efficient, production-ready SLMs — including Phi-3, TinyLlama, Gemma-2B, Qwen2-0.5B, and Mistral-7B variants — optimized for fine-tuning via LoRA, QLoRA, and IA³. Support extends to custom architectures via ONNX-compatible exporters, enabling seamless integration with specialized inference runtimes like llama.cpp or Ollama.
Can I bring my own training data format?
Absolutely. Radal accepts CSV, JSONL, Parquet, and ZIP archives containing text, instruction pairs, or chat logs. It auto-detects schema, suggests preprocessing (e.g., deduplication, truncation, template alignment), and enables visual validation — previewing how your data maps to training examples before a single epoch begins.
How does the AI Copilot improve model quality?
The Copilot doesn’t just suggest parameters — it reasons across your entire stack: dataset statistics, hardware constraints, latency budgets, and evaluation goals. If your test set shows bias toward certain entity types, it recommends targeted data augmentation. If inference is slow on ARM64, it proposes GGUF quantization + KV-cache optimizations — all surfaced as editable visual nodes, not opaque commands.
Do I need to manage infrastructure to use Radal?
No. Radal offers fully managed cloud training with elastic GPU scaling, or self-hosted deployment options (Kubernetes, Docker Compose). You retain full control over compute location, network policies, and data residency — with no vendor lock-in on model formats or training artifacts.
Where can I get started today?
Visit radal.ai to sign up for a free tier with unlimited visual workflows and 5 hours of managed training compute per month. Explore interactive demos at app.radal.ai, request an enterprise pilot, or schedule a technical deep-dive with our AI Solutions team — all without requiring a credit card or sales call.