Adaption Labs Introduces 'Invent a Dataset': Training Data Generated From a Task Description, Not a Seed Corpus Adaption Labs推出「 Invent a Dataset 」：从任务描述而非种子语料库生成训练数据

TL;DR

• Adaption Labs released "Invent a Dataset," a feature that generates structured, training-ready datasets from natural language behavior descriptions without requiring a seed corpus, predefined schema, or labeling guide
• The API uses domain codes (e.g., medical, medical.symptoms_diagnosis) as primary controls, with async generation via datasets.invent and datasets.get polling
• Two output formats are supported: instruction_dataset (prompt-completion pairs for SFT) and preference_pairs (chosen/rejected pairs for DPO-style training)
• Language expansion supports both translate and localize modes, with sample_rate control and credit billing on expanded output rows
• Dataset IDs integrate directly with AutoScientist (launched May 2026), which co-optimizes data and training recipes, reportedly beating human-configured training by 35% across eight verticals with win rates improving from 48% to 64%

Why It Matters

This represents a significant shift in the synthetic data pipeline by eliminating the need for existing seed data or manual schema design, allowing practitioners to generate training data directly from behavioral intent. The tight integration between data generation and AutoScientist's co-optimized training creates a closed-loop system that could substantially reduce the time and expertise required to produce domain-specialized models. For organizations working with proprietary or specialized tasks where relevant data is scattered across internal systems, this approach bypasses the traditional bottleneck of data collection and labeling.

Technical Details

• API Mechanics: A single datasets.invent call initiates async generation; status is polled via datasets.get until succeeded or failed. The estimate=True parameter allows cost previewing without charges. An idempotency_key (up to 255 characters) ensures safe network retries.
• Domain Control: Domain codes are fetched dynamically via datasets.invent_domains rather than hardcoded. Subdomain narrowing (e.g., medical.symptoms_diagnosis) is supported, and multiple domains can be combined in a single run.
• Output Formats & Ports: Data downloads as JSONL, JSON, CSV, or Parquet. The instruction_dataset format produces SFT-ready prompt-completion pairs, while preference_pairs generates chosen/rejected pairs for preference optimization like DPO.
• Language Expansion: Two modes—translate (language variants) and localize (country-language pairs with locale-specific wording). A sample_rate parameter (0.01–1) controls expansion fraction, with credits billed on expanded row count.
• AutoScientist Integration: Dataset IDs feed directly into autoscientist.create, which co-optimizes data and training recipes. Evaluated across eight verticals with 5,000–100,000 row datasets on Together AI fine-tuning architectures, showing 35% average improvement and win rate gains from 48% to 64% versus staff-configured training.

Industry Insight

The "zero-data loop" paradigm—generating data from intent and immediately feeding it into co-optimized training—could become a standard workflow for vertical AI development, particularly in regulated or specialized domains where labeled data is scarce or proprietary. Practitioners should evaluate whether the hosted-only generation model (no self-hosted path documented) aligns with data governance requirements, especially for sensitive domains like healthcare. The reported 35% improvement over human-configured training suggests that automated data-train co-optimization may soon outperform traditional manual pipeline design, making it worth benchmarking against existing fine-tuning workflows before committing to custom data collection efforts.

TL;DR

• Adaption Labs推出"Invent a Dataset"功能，可从行为描述自动生成结构化训练数据集，无需种子语料、预定义模式或标注指南
• 支持instruction_dataset（SFT）和preference_pairs（DPO）两种输出格式，可导出为JSONL/JSON/CSV/Parquet
• 与AutoScientist训练平台打通，形成"意图→训练模型"的零数据闭环，内部评估显示性能提升35%
• 支持多领域代码控制、多语言/本地化扩展，通过Python SDK和REST API调用

为什么值得看

这篇文章介绍了一种全新的训练数据生成范式——从"行为描述"直接生成训练数据，跳过了传统数据工作中最耗时的标注和清洗环节。对AI从业者而言，这意味着可以更快迭代专有领域模型，降低数据准备门槛。

技术解析

• 核心机制：通过datasets.invent API调用，传入领域代码（如medical、medical.symptoms_diagnosis）和行为描述prompt，异步生成训练行。支持estimate=True预览费用、idempotency_key保证幂等性。
• 输出格式：默认instruction_dataset生成prompt-completion对用于SFT；preference_pairs生成chosen/rejected对用于DPO等偏好训练。
• 语言扩展：language_expansion支持translate（多语言变体）和localize（国家+语言本地化），通过sample_rate控制扩展比例。
• 闭环集成：生成的dataset ID可直接传入autoscientist.create，实现数据生成→训练→评估的自动化流水线。

行业启示

• 数据工程范式转变：从"先有数据再训练"转向"按需生成数据再训练"，可能重塑小样本/专有领域模型开发流程。
• 合成数据质量验证：需关注生成数据的真实性、偏差控制和下游任务表现，建议建立严格的评估基准。
• 平台锁定风险：生成运行在Adaption托管平台，无自托管方案，企业需权衡便利性与数据主权。