Compute
Managed cloud compute for Python, AI, and machine learning jobs.
Start with an outcome, such as recognizing your dog in photos, or bring a specific model and training script. Your coding agent can use Orchard to recommend public model candidates, plan the experiment, match it to available compute, run it, and review the measured results.
Orchard supports both recommended and builder-specified workflows. Choosing your own model is a preference, not a skill-level label, and both paths can use the same estimate, execution, and validation tools.
How Orchard responds
If your request already names a model, code, training settings, or method, your agent should follow those details and go directly to a compute estimate. If your request is open-ended, it should ask whether you want Orchard to recommend a model and training method, or validate model or code you already have.
- Recommended path: Orchard searches for suitable public models and proposes a training and evaluation plan.
- Builder-specified path: you choose the model, code, and settings; your agent reviews compatibility, then Orchard estimates and executes the run you approved.
- Either path: you can record a target metric, generate a standard report, review weak areas, and plan the next experiment.
Before model selection
A useful recommendation needs more than a broad idea. Your agent may inspect the project and ask for the missing details that change the model or evaluation plan:
- What the model should predict, classify, detect, or generate
- Which field, label, or example counts as the correct answer
- The primary metric and the target value you care about
- Latency, runtime, memory, and cost constraints
- Whether the model and data must allow commercial use
Model discovery and confirmation
Orchard currently searches the public Hugging Face model catalog. It filters and ranks candidates using task fit, declared license, access, publisher, file format, model size, and public adoption signals. Downloads and likes can help compare candidates, but they do not prove quality for your data.
Your agent must show the candidates and their tradeoffs. Orchard will not connect a recommended model to a paid training run until you confirm that model. You may instead provide a specific model or code from another source after its license and compatibility are reviewed.
Estimate and approval
Every run begins with current Cheapest, Balanced, and Fastest options. The estimate describes the machine, expected duration, hourly price, and estimated total. Availability changes, so an estimate is short-lived.
Model confirmation and compute approval are separate decisions. After you choose an option and approve a maximum cost, Orchard starts the job and stops it automatically when it completes or reaches the approved cost or runtime limit. If the current price rises above the approved maximum, your agent must estimate again and ask before starting.
Training and feedback loop
Before the job starts, your agent should define a fixed evaluation split, primary metric, metric direction, and optional target value. Training code can write Orchard's standard report with real train, validation, and test metrics, data-quality counts, and performance for important data slices.
After the run, Orchard validates that report and can identify a train-to-evaluation gap, weaker slices, missing labels, duplicates, and whether an untouched test result reached the recorded target. It then proposes one evidence-based next experiment. Your approval is required before changing the data, model, training settings, evaluation rules, or starting another paid run.
If the standard report is missing, Orchard returns the logs and outputs but does not invent metrics or claim that the model is good.
Ask your agent
Paste one of these into your coding agent:
- "Train a model that recognizes my dog in photos. Recommend the model and evaluation plan first."
- "Fine-tune this image classifier and target at least 90% macro F1 on the test set."
- "Use the model and training script already in this project. Estimate compute before running it."
- "Find a commercially usable public model for text classification, then show me the candidates."
- "Review the completed training run and suggest the next experiment based on its report."
- "Backtest this Bitcoin strategy with time-ordered splits, fees, slippage, and a maximum drawdown metric before any live integration."
Time series, trading, and other advanced work
Orchard can help research and execute LLM fine-tuning, generation, speech, recommendation, time-series, reinforcement-learning, and trading experiments when suitable code and evaluation rules can be prepared. These tasks usually need more clarification and stronger validation than a standard classifier.
Trading work should first define the prediction or policy target, use time-ordered and leakage-safe splits, include fees and slippage, compare simple baselines, and run backtests before any live connection. Orchard can automate experiments; it does not promise profit or remove the need for risk review.
Results and current limits
Your agent can monitor job status and recent logs, stop a job early when you request it, and download the result archive after completion. The archive contains the requested output files, run log, exit code, and the training report when one was produced.
- Compute runs uploaded project archives and command-line Python or ML workflows; it is not an interactive notebook.
- Do not place passwords, tokens, or other secrets in the project archive or run command.
- Model recommendations are starting points, not guarantees of accuracy, safety, fitness, or license suitability.
- Final release decisions should also review production-like data, latency, cost, safety, and licensing.
Compute pricing
You can discover models and compare compute estimates on any plan. Starting a cloud compute job requires an active Builder plan, an account spending cap, and explicit approval of the maximum cost.