Skip to main content
Fine-tune an open model on your own production traffic. Belvedir takes the successful tasks it segmented from your traces, distills them into a training set, and trains a LoRA adapter on a GPU sandbox. The finished adapter appears on the Models page and is yours to download.

​
What it trains on

Successful tasks that captured both an input and an output — shown as the example count on the Training page. You can train on everything the agent does well, or pick a single group to distill one kind of work into a specialist. Training needs at least 200 examples; keep sending traces until you cross the floor. Datasets are capped at 1,000 examples — the newest eligible tasks win.

​
Training methods

Three methods, all consuming the same success-only dataset:
  • Supervised fine-tuning ($10.00) distills successful tasks directly — the model learns to reproduce the outputs your agent got right; the default and the cheapest.
  • On-policy self-distillation ($20.00) has the model generate its own answers and nudges them toward what a teacher — the same base model shown the known-good output — would say, so it learns from its own mistakes instead of copying.
  • Reinforcement learning ($30.00) scores sampled completions — with a Claude judge or by similarity to your successful outputs — and trains to raise that score (GRPO or PPO, picked automatically).

​
Base models

Small open models that fit a single GPU: Qwen 2.5 7B / 1.5B Instruct, and Llama 3.1 8B Instruct. Llama is gated on HuggingFace, so it asks for an HF token — the token is passed to the training sandbox only and never stored.

​
Runs are metered

Each training run is charged to your balance when it starts (the method prices above) and refunded automatically if the sandbox fails to launch. Runs are capped at 2 hours; typical runs finish well under that.

​
Auto-training

Instead of launching runs by hand, enable auto-training on the Training page. It triggers a new run either when enough new traces have arrived since the last run (threshold of 50 or more) or on a fixed cadence (every 1–365 days), with a 24-hour cooldown between runs. Auto runs repeat the configuration of your last completed run and are billed the same as manual ones.

​
What you get back

Training and eval loss on a held-out split, recorded on the run, plus the LoRA adapter itself (a zip of the standard PEFT files). Load it on the base model with peft, vLLM’s --lora-modules, or any runtime that speaks LoRA. Always-on hosted serving from Belvedir is on the roadmap.

​
Run inference

Every finished model has a Run inference button on the Models page. It boots a GPU sandbox serving the base model with your adapter attached and opens a chat panel against it, so you can try the model before wiring it into anything. A session costs $2.00, takes a few minutes to warm up (the sandbox downloads the base weights), and shuts down automatically after 30 minutes.