Soramai
Sign inGet started

Soramai · Docs

Getting started.

Sign up, top up, prepare a dataset, fine-tune a LoRA adapter, test it in the playground, and deploy it as a live API. Total time: about 15 minutes the first time, 5 minutes after that.

Open dashboardAll docs

Before you start

One account, one supported browser, and a few dollars of credits is all you need. No CUDA installs, no GPU drivers, no Kubernetes.

  • A Soramai account. Created via Google or Apple sign-in. New accounts receive 30 starter coins so you can run a small fine-tuning job before topping up.
  • A modern desktop browser. Chrome, Safari, Firefox, or Edge from the last two years. Mobile works for monitoring existing runs but the fine-tuning launch flow is desktop-first.
  • A dataset (or a prompt). Either bring your own JSONL file, or describe what you want and let Soramai generate it for you.
  • $5–$25 in credits for your first end-to-end run. See the pricing page for exact pack sizes.

Step 1 · Create your account

Soramai uses passwordless sign-in. Apple or Google handles the credential and recovery.

  1. 1Go to soramai.com/signup.
  2. 2Click Continue with Google or Continue with Apple.
  3. 3Approve the OAuth consent screen. Your account is created instantly with 30 starter coins.
  4. 4You’ll land on the dashboard with your wallet balance visible top-right.

Step 2 · Top up credits

Skip this step if 30 starter coins is enough for your first run. A 100-step Beginner-tier fine-tuning costs about 8 coins.

  1. 1On the dashboard, scroll to Add credits.
  2. 2Pick a pack. The 20,000 coins / $24 pack is the most common starting point — it covers 30–60 small fine-tuning runs.
  3. 3Click Checkout. You’re redirected to Stripe.
  4. 4After payment your balance updates within 30 seconds. A receipt is emailed automatically.

Credits never expire while the account is active. Refunds for failed jobs caused by platform faults (worker crash, GPU eviction, adapter validation failure) are issued automatically — you only pay for fine-tuning time that produced a usable result.

Step 3 · Prepare a dataset

You have three options. Pick whichever matches what you already have on hand.

Generate with AI →

Describe the data you want and let Soramai produce it. Great when you want a fast prototype without curating examples by hand. 500-row datasets generate in 60–120 seconds.

Write it yourself →

Open Dataset Studio in the dashboard and type rows directly. Best when you have specific examples in mind and want full control.

Upload a JSONL →

Drop a .jsonl file into Dataset Studio. Each line is a JSON object with prompt and response fields. Up to 20 MB / 50,000 rows.

Minimum viable dataset shape (JSONL):

{"prompt":"Summarise the changelog: ...","response":"Released 1.4 ..."}
{"prompt":"Classify this ticket as bug/feature: ...","response":"feature"}
{"prompt":"...","response":"..."}

Both prompt and response are required. Optional system_prompt is honoured if present.

Step 4 · Launch a fine-tuning run

Two clicks once your dataset is ready.

  1. 1Open soramai.com/train.
  2. 2Pick a base model from the dropdown. Beginner (Qwen 2.5 7B) is the default and the cheapest. Switch to Pro or Elite tiers for larger models.
  3. 3Click Pick from My Datasets and select the one you prepared. Or paste a signed URL.
  4. 4Pick a step count. 100 steps is a good first run — ~3 minutes, ~8 coins. Bump to 500 once you’re happy with the recipe.
  5. 5Confirm the cost estimate, click Launch fine-tuning. You’ll be sent to the live fine-tuning detail page.

The dashboard live-bills the run per second. You can safely close the tab — the worker keeps running and your balance updates in the background. A push notification arrives when the run completes (if you have the iOS app installed).

Step 5 · Test in the playground

Verify the model learned what you wanted before deploying. The playground is free except for GPU runtime.

  1. 1From the completed fine-tuning page, click Open in playground.
  2. 2A warm inference pod starts (~30 seconds first time). After that, messages return in under a second.
  3. 3Try 5–10 prompts that represent your real workload. If the model nails them, you’re ready to deploy. If not, add more fine-tuning examples and re-train.
  4. 4Playground sessions auto-terminate after 10 minutes of idle to save coins. You’re billed per minute of warm GPU time.

Step 6 · Deploy as an API

One click to turn a fine-tuned adapter into an autoscaling HTTPS endpoint your code can call.

  1. 1From the completed fine-tuning page, click Deploy. Or open soramai.com/deployments and pick from your model list.
  2. 2Set a deployment name (just for your reference) and confirm.
  3. 3Soramai provisions a serverless inference endpoint. The dashboard shows the URL and an API key (shown once — copy it now).
  4. 4Test the endpoint with the curl snippet shown on the deployment page, or jump to the Inference & Deploy guide for the request schema.

First request:

curl https://soramai.com/api/v1/inference \
  -H "Authorization: Bearer sk-ai-xxxxxxxxxx" \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Hello"}'

Next steps

You’ve done the full loop. From here, pick whichever path matters most for your project.

Datasets reference →

JSONL schema, image-dataset format, validation rules, multi-dataset merging, AI generation parameters.

Fine-tuning reference →

Base model catalogue, LoRA hyperparameters, monitoring metrics, refund policy for platform faults.

Inference & Deploy →

Playground vs Deploy API, request and response shapes, streaming, rate limits, billing.

Pricing →

Coin pricing, per-tier rates, worked examples for typical fine-tuning and inference workloads.

Security →

Data handling, isolation, encryption, responsible disclosure.

Status →

Live service health for the dashboard, fine-tuning pool, inference endpoints, and billing.