ZeqouXTraining
Train. Tune. Test. Build.
A local-first training studio. Pick a model and a dataset, choose LoRA, QLoRA, SFT, a full fine-tune — or train a small transformer from scratch with nothing but Python. Runs are real processes with real gradients, checkpoints you can resume, and a lineage tree you can inspect. Models, datasets and logs are ordinary files in a workspace folder you choose; nothing is uploaded anywhere.
zeqouxtraining — run qwen2.5-0.5b-lora- Step
- 120 / 300
- Loss
- 1.284
- VRAM
- 9.1 / 12 GB
The XTraining workspace, previewed
Model → Dataset → Method → Settings → Train
One Training page takes you through it: pick a model from the Hugging Face Hub or a local folder, point at a dataset, then choose LoRA, QLoRA, SFT, a full fine-tune — or train a small transformer from scratch with no base model at all. Simple mode shows the settings that decide whether a run succeeds; Advanced mode hands you the rest — quantization, scheduler, LoRA rank, checkpoint retention, the scratch architecture. Presets fill the whole form in one click, so a first run is two minutes of reading, not an afternoon of YAML.
It checks before it starts
Real hardware detection through the CPU, the driver, nvidia-smi and torch.cuda; a memory estimate for the run you configured, labelled as an estimate; and dataset validation that reports field mapping, duplicates, empty rows and over-length samples. Every blocker is collected in one place — a missing runtime, a dataset error, an estimate that will not fit — before you press start.
Works before you install anything — and never fakes it
Two backends ship with the app. The from-scratch backend is pure Python: it needs nothing but an interpreter and trains as soon as you press start. The Hugging Face backend lights up when PyTorch, transformers and peft are present — and PyTorch has no wheels for some Python versions, which the app says out loud instead of failing later. The ML runtime is never bundled and never installed behind your back: it is machine-specific and several gigabytes, so the app discovers the interpreters you already have, explains which one can run it, and installs only when you ask — into its own environment inside the workspace, never into your system Python.
Watch the run, keep the result
Live loss and learning-rate charts, logs and checkpoints while it trains. Stop and pause are cooperative, so a paused run resumes from its last checkpoint — weights, optimizer, scheduler and RNG restored, with the engine reporting per item what was really restored. Continuing a finished model is a separate, deliberate operation with a fresh optimizer, and every model keeps a parent link, so the Experiments page can draw the real lineage tree. When it finishes, evaluate it, test it in the built-in playground, serve it locally, or merge and export the adapter with a model card.
Get XTraining
Free for Windows, macOS, Linux. Needs Python 3.10 – 3.13; the from-scratch backend needs nothing else, and PyTorch only becomes necessary when you fine-tune a real Hugging Face checkpoint. Installed builds update themselves from the release channel.