training & evaluation
Make a model yours
Teach a model your codebase, your tone or your task — on this Mac, from data you can see — and prove the result with held-out evaluation before you use it.
Fine-tuning on Apple Silicon
The Train tab in the model library runs jobs in a local MLX worker and shows them live. A memory profiler estimates what a job needs before it starts.
- LoRA, DoRA and QDoRAAdapter training with quantized base models, or full fine-tuning when memory allows.
- Optimizer choiceAdam, AdamW, SGD and Muon, with gradient checkpointing to trade compute for memory.
- Preference alignmentDPO on chosen and rejected pairs, and reinforcement learning with GRPO.
- DistillationSequence, logit and on-policy distillation from a teacher you select — a local model or an endpoint you run.
- Exact continuationCheckpoints save optimizer, scheduler, random state and dataset position, so a stopped job resumes exactly.
- Live curvesLoss and gradient-norm charts while training, to catch overfitting and divergence early.
- Beyond textVision-language adapters, speech and music adapters, and image-diffusion LoRAs from licensed captions.
- Prune and pretrainStructured pruning with quality gates, and small model or tokenizer experiments on licensed corpora.
- ExportFuse adapters, quantize and export GGUF for any runtime, or hot-swap adapters in the Orvlet Engine.
A trained model is only as good as the test it never saw.
Photograph — Benjamin Voros via Unsplash
Data you can trust
Most of a good model is a good dataset. Orvlet keeps every record's provenance in view.
Find datasets
Browse Hugging Face and ModelScope datasets, preview rows and schema, and download revision-pinned files only when you ask.
Map your own files
Turn CSV, TSV, JSON and Parquet into training records by mapping columns, with row-level validation.
No leaks
Assemble sources with explicit train, validation and test splits, deduplicate normalized inputs across them, and keep a source-hash manifest.
Synthetic data, reviewed
Save generation recipes, create bounded samples with a model you choose, and review provenance before accepting a record. Generated examples are admitted only when their verifier actually ran and passed.
Preference pairs
Write or import prompt, chosen and rejected pairs, review each one, and export accepted pairs for alignment.
Evaluation that holds up
Compare models and changes honestly — on your tasks, not a leaderboard's.
- Blind arenaTwo models answer the same prompt anonymously; you vote, then see who was who, with Elo-style ratings across tasks.
- Side-by-side chatsPersistent comparisons with synchronized input and independent histories, saved as reusable presets.
- Model judgesConfigured judges score completed runs with thresholds, sampling and classifications.
- Trace evaluationGrade whole agent trajectories with rubrics and deterministic graders, and compare prompt candidates.
- Training reportsHeld-out results tied to the evaluation split, base model, adapter and the reloaded exported artifact — with missing measurements shown as missing.
- Plugin and skill evalsRun a plugin or skill against cases with and without it, and see per-case differences and cost.
- BenchmarksAgentic capability and reasoning benchmark runners for local models, plus prompt-processing and generation speed tests.
- Quantization trade-offsMemory, speed and perplexity across quant levels, side by side.
- Classical MLCreate, train, evaluate and export classification, regression and forecasting models from local tabular and time-series data.
Bigger jobs, your hardware
When a job outgrows the Mac, submit it to a machine you own over SSH. Orvlet shows exactly which files go up — names, sizes and checksums — monitors the run, and recovers the artifacts when it finishes or is cancelled.
Start small
A first adapter on a small local model is a short job on a recent Mac. Pick a model that fits comfortably, start with a few hundred examples, and let the evaluation report tell you whether it helped.