HTTP API

The Python SDK is a thin client over three HTTPS endpoints: one per model, and one that chains them. Anything the SDK does, you can do with a POST.

Endpoint	Does
`POST /audit`	Runs `sdvm-audit-1`; returns each sample unchanged plus an `audit` block.
`POST /fix`	Runs `sdvm-fix-1`; returns each sample plus a `fix` block.
`POST /refine`	Runs `sdvm-audit-1 -> sdvm-fix-1 -> sdvm-audit-1`; returns each sample with `audit` (before the fix), `fix` and `reaudit` (after).

Each endpoint takes an optional config for the stages it runs: {"votes"} on /audit, {"max_attempts"} on /fix, and {"audit": {...}, "fix": {...}, "reaudit": {...}} on /refine. See Core concepts.

Base URL

https://api.sdvm.ai

Authentication

Send your API key as a bearer token. Keys are created on your profile page.

curl https://api.sdvm.ai/audit \
  -H "Authorization: Bearer $SDVM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "data": [
      {
        "task_type": "multiple_choice",
        "context": "What is the capital of France?",
        "choices": ["Berlin", "Paris", "Madrid", "Rome"],
        "answer_index": 1,
        "style": "qa"
      }
    ],
    "config": {"votes": 3}
  }'

The sample shape

Every element of data is a JSON object with a task_type. Unknown keys on a multiple-choice sample are passthrough metadata and come back untouched.

text
multiple_choice
question_answer
conversation

{ "task_type": "text", "text": "she sels sea shells by the sea shor" }

{
  "task_type": "multiple_choice",
  "context": "A woman applies mascara. she",
  "choices": ["applies it to her lashes.", "drives off.", "mixes a bowl.", "exits."],
  "answer_index": 0,
  "style": "continuation"
}

style is "continuation" for a sentence stem (HellaSwag) or "qa" for a complete question (MMLU). See Multiple choice.

{
  "task_type": "question_answer",
  "question": "Natalia sold 48 clips in April and half as many in May. How many altogether?",
  "answer": "She sold 48 / 2 = 24 clips in May, so 48 + 24 = 72 altogether."
}

Two fields, no choices. answer is whatever the dataset publishes — a bare value, a sentence, or a worked solution. See Question and answer.

{
  "task_type": "conversation",
  "messages": [
    { "role": "user", "content": "how do i reverse a list in python" },
    { "role": "assistant", "content": "Use slicing: `xs[::-1]` returns a reversed copy, or `xs.reverse()` in place." }
  ]
}

A chat transcript in the OpenAI messages shape, roles system / user / assistant. turns ("last" default, or "all") is how much of it the models read. See Conversation.

Responses carry the samples back under audited_data, fixed_data or refined_data, with counts, cost_usd, input_tokens and output_tokens.

Errors

Status	Meaning
401	Invalid or revoked API key.
402	Insufficient credits.
422	Validation error: empty or oversized list, malformed sample, sample too large for the model, or estimated cost over the per-request limit. The `detail` field says which.
429	More than 100 requests per minute.

Limits

Up to 100 samples per request and 100 requests per minute per key. Pricing is token-based with a $0.01 minimum per request; see Pricing.

The endpoint pages in this section are generated from the service’s OpenAPI specification, so field names, defaults and ranges match what the server enforces.

Get started

Sample types

Models

Guides

SDK

Pricing and limits

HTTP API

Base URL

Authentication

The sample shape

Errors

Limits

​Base URL

​Authentication

​The sample shape

​Errors

​Limits

Base URL

Authentication

The sample shape

Errors

Limits