Agent training data

Training data you can verify before you trust.

Spring of Winds Labs is a small research lab in Japan, founded in 2026. We build agent training data — and we publish the measurements to prove it.

Spring of Winds Labs — verifiable agent training data

Proof

Measured, not claimed

28,922validated rows (every shipped row, no sampling)
100%schema pass, last measured 2026-11-19
Public validatorstdlib-only; run it on your own copy
1,300+downloads on Hugging Face (Nov 2026)

Per-lot results: EVALS.md · Details and a verbatim row from a shipped lot: evidence page.

Measured pilot — $990

Know what the data does on your task before you license it

You give us your tasks. We measure your model with and without our data, and deliver the report. The price is fixed at $990.

  1. Your task. You define a set of tasks that matter to your agent.
  2. With / without. We measure the same tasks on the base model and on the model trained with our data.
  3. The report. You receive a written report with the three metrics below.
MetricWhat it tells you
Recovery success rateHow often the agent recovers after an injected tool failure
Schema & validation error rateHow often outputs break the declared tool schema
Task completion rateHow often the task is completed end to end

Pilot credit: 100% of the pilot fee is credited if you license within 60 days of the report; 50% after that.

MCP failure recovery in a measured pilot
Example pilot report

Our own rehearsal, Run 1: 1.7% → 56.7% recovery

The base model alone recovered from 1.7% of injected MCP failures. With a LoRA adapter trained on our own corpus, it recovered from 56.7%.

Metric (60 tasks)A: base model onlyB: base + LoRA
Recovery success rate1/60 = 1.7%34/60 = 56.7%
Task completion rate1/60 = 1.7%31/60 = 51.7%
Schema / validation error rate0/60 = 0.0%1/60 = 1.7%
Measured on our own fixed task set — your pilot runs on yours. This is a seller-run rehearsal, not a customer result. Conditions: 60 MCP failure-recovery tasks, base model Qwen/Qwen2.5-1.5B-Instruct, LoRA trained on 2,000 rows from our existing corpus, measured 2026-11-19. Recovery was weakest on rate-limit tasks (3/10 with LoRA), and the LoRA run produced 1 schema error where the base model produced none.

Read the full report →

Ask about a pilot on your task

Catalogue

Four tracks, one schema discipline

Every track ships as Strict JSONL. Prices are for the 1K lot: perpetual, organization-wide licence, no redistribution. Free 50-row samples (Apache-2.0) are on Hugging Face.

MCP Agent Trajectory Strict JSONL

MCP Agent Trajectory

  • Multi-turn MCP tool-use trajectories
  • 403 / 429 / lock-conflict failure and recovery
  • 1K lot: 1,003 rows · Extended: 3,303 rows (1K lot + 2,300 non-overlapping)

$7,300 · 1K lot

Acquire licence

Free sample on Hugging Face
Function Calling EN Strict JSONL

Function Calling (EN)

  • Argument-level correctness under changing state
  • Declared tools and argument keys only
  • 1K lot: 1,000 rows · Extended: 3,300 rows (1K lot + 2,300 non-overlapping)

$6,400 · 1K lot

Acquire licence

Free sample on Hugging Face
Function Calling JA Strict JSONL

Function Calling (JA)

  • Single-shot Japanese request → strict-JSON tool call
  • Honorifics and entity ambiguity coverage
  • 1K lot: 1,000 rows · Extended: 3,300 rows (1K lot + 2,300 non-overlapping)

$8,500 · 1K lot

Acquire licence

Free sample on Hugging Face
Regulatory Compliance CoT Strict JSONL

Regulatory Compliance CoT

  • Multi-jurisdictional audit reasoning
  • Ordered reasoning steps, grounded conclusions
  • 1K lot: 1,000 rows · Extended: 3,300 rows (1K lot + 2,300 non-overlapping)

$12,500 · 1K lot

Acquire licence

Free sample on Hugging Face

Starter 250 — $49

250 MCP error-recovery trajectories — a strict subset of the MCP 1K lot, same schema, same validation. Instant download.

Get Starter 250

How it's made

Synthetic by design, validated row by row

Generation and validation pipeline

Ways to work with us

Three doors, from lowest to highest commitment

Starter 250 — $49

Instant download on Gumroad. Look at real rows in your own pipeline.

Get Starter 250

Measured pilot — $990

Fixed price. With/without measurement on your task, delivered as a report.

Ask about a pilot

Licence

Direct purchase with Stripe invoicing for the 1K and Extended lots.

Request a licence invoice

Evaluation or custom specification production: write to springofwindslabs@gmail.com.

About the lab

Small and new, so everything can be checked first

We are small and new. That is exactly why everything we sell can be checked first: samples, schemas, validation, and a measured pilot on your own task before any licence.

Contact: springofwindslabs@gmail.com

FAQ

Short answers

What does the licence cover?

Perpetual, organization-wide use for model training, evaluation and internal tooling.

Can I redistribute the data?

No. There is no redistribution tier. Free 50-row samples are Apache-2.0.

Do early buyers get updates?

Founding buyers receive free updates within the same track.

Do you offer exclusive licences?

No. We do not sell exclusive licences.