Adaption Labs Generates Training Data From a Task Description Alone

Adaption Labs shipped Invent a Dataset on 3 September 2026, a capability that generates structured, training-ready rows from a plain-language description of the behaviour a model should learn — no seed corpus, no predefined schema, and no labelling guide required. The release is live in the Adaption app and through both the Python SDK and a one-to-one REST API mirror. For teams stuck at the front end of the post-training pipeline, the constraint being removed is not compute or annotation throughput; it is the assumption that useful training signal already exists somewhere in an internal corpus.

The framing targets a real ceiling. When available data only approximately matches a specialised task, fine-tuning quality is bounded by that approximation. Proprietary workflows, clinical edge cases, and domain-specific reasoning rarely convert cleanly into focused instruction sets — a problem that, as we have covered in pipeline architecture, not better models, as the source of AI gains, often goes unsolved because teams default to scaling compute rather than fixing data provenance.

API Mechanics and Control Surface

A single call to datasets.invent starts generation and returns immediately with status: running. Callers poll datasets.get until the status resolves to succeeded or failed, then download the artifact. Output formats are JSONL, JSON, CSV, and Parquet — portable files the caller owns and can use with any training stack.

Domain targeting is the primary steering mechanism. Valid codes come from datasets.invent_domains, not from hardcoded strings, so the set can expand as Adaption adds coverage. At least one domain or subdomain is required; passing a top-level domain such as medical without a subdomain draws from its full scope, while a qualified subdomain like medical.symptoms_diagnosis narrows generation to that slice. Multiple domains combine within a single run.

Three production-oriented parameters complete the surface. estimate=True calculates the credit cost of the exact request and returns estimated versus available credits without creating a dataset or charging anything. prompt accepts up to 10,000 characters to steer row content. idempotency_key accepts up to 255 characters and causes retries to return the original dataset rather than launch a duplicate run. Row counts per launch are capped by plan tier.

Two training_type values are supported. instruction_dataset (the default) produces prompt-completion pairs suited to supervised fine-tuning. preference_pairs produces chosen and rejected completion pairs for preference-optimisation methods such as DPO.

language_expansion runs in two modes. translate generates a new row variant per target language. localize generates a variant per country-and-language pair, using locale-appropriate wording rather than direct translation. A sample_rate between 0.01 and 1.0 controls what fraction of invented rows receives expansion; billing applies to the expanded row count, not the source count. Unsupported codes return a 400 with a sample of valid values.

The AutoScientist Loop

Invent a Dataset is the first half of a closed pipeline. The generated dataset ID passes directly to autoscientist.create, which co-optimises the data and the training recipe against the stated objective. AutoScientist launched in May 2026 and is the training-side counterpart to the Adaptive Data pillar.

Adaption's reported figures are specific: an average 35% gain over configurations built by the company's own research staff, across in-house domain-specialised evaluations spanning eight verticals. Win rates moved from 48% to 64%. Dataset sizes in those evaluations ranged from 5,000 to 100,000 rows, on architectures offered for fine-tuning by Together AI. These are internal evaluations, not third-party benchmarks — a limitation worth weighing against findings from more independently reproducible settings.

Generation runs entirely on Adaption's hosted platform and consumes credits. No self-hosted generation path is documented.

The broader pattern is that the data-generation layer is consolidating around intent-driven APIs rather than corpus-manipulation tooling. When a single call can move from a behavioural description to a preference-pair dataset without any intermediate human-defined schema, the post-training bottleneck shifts further toward evaluation and deployment — pressures that infrastructure governance addresses from a different angle for the same downstream moment.