Alrithm
Data API · v1 live

Verified reasoning data.
One API call away.

Point your training tool at a single URL. Every row arrives already formatted for fine-tuning, with cryptographic proof that the answer is correct.

Get startedRead the docs →
arithm-api — terminal
AnnouncementOct 4, 2026

The catalog is now organized by task type

Every dataset is now tagged as solve, debug, optimize, explain, or review.

Read more

Available now

Verified reasoning datasets, streaming as newline-delimited JSON. One API call, no SDK.

View all →
refusal_defer

Refusals: Legal and Licensing Questions

Requests that need a licensing or legal determination and cannot be computed; the correct response routes them to the authority.

200 rows550.5 KB
refusal_infeasible

Refusals: Infeasible Goal

Well-posed requests whose goal has no solution; the correct response says so instead of inventing a number.

200 rows574.1 KB
refusal_inconsistent_input

Refusals: Inconsistent Input

Requests whose individually valid values contradict each other; the correct response asks to reconcile them.

200 rows583.3 KB
refusal_invalid_input

Refusals: Invalid Input

Requests where a single value lies outside its valid range; the correct response asks for a corrected value.

200 rows544.9 KB
refusal_underspecified

Refusals: Underspecified Requests

Requests with several valid readings whose answers differ; the correct response asks which reading is meant.

200 rows645.6 KB
refusal_missing_parameter

Refusals: Missing Parameter

Requests where one computation fits but a required field is absent; the correct response names the missing field.

200 rows505.2 KB
Get started
16,800
Verified rows
84
Published datasets
84.7 MB
Corpus size
SHA-256
Provenance hash

What a row looks like

Each row carries the problem, a reference implementation, a step-by-step derivation, and the exact output of the algorithm. Re-run the solution and you get the same answer, byte for byte.

Rows stream as newline-delimited JSON. Point any tool at the URL and read them line by line.

GET /api/v1/models/alr-code-platform/datasample.jsonl
{
  "sample_id": "a49939cd-a74e-4172-a4db-6f45eaf4f1be",
  "sample_type": "debug",
  "category": "dynamic_programming",
  "difficulty": "medium",
  "question_text": "Find the length of the longest common subsequence…",
  "headline_answer": 3,
  "provenance_hash": "222289f0fa0e933f87d875d3e32f1a66…"
}

How it works

01
Sign up

Create an account in under a minute, then generate your first API key from the dashboard.

02
Verify

Every row carries a reference solution and a hash. Re-run the solution yourself to confirm the answer.

03
Train

Rows arrive in OpenAI fine-tune format. Pipe them straight into a training job.