Adaption reports that two generative AI models fine-tuned on examples from its Invent API produced medical-question answers that Google’s Gemini 3.1 Pro ranked first more often; a third model did better without fine-tuning. The company’s Oct. 1 technical report compares Invent with five model APIs for generating the examples, then trains models on the resulting datasets.
Invent starts from a description of the behavior a team wants a model to learn and returns structured training examples. “Zero data” here means no examples already exist for that task; the user still has to describe it. Adaption introduced the product in September. The new report tests whether its generated examples hold up against data created with other language-model APIs.
The researchers tried eight dataset requests spanning different tasks, domains and languages, with up to 20,000 examples per request. At 5,000 examples, Invent received a rubric-based quality score of 7.90 out of 10, compared with 6.76 for Claude Opus 5 and 6.73 for GLM-5.3. Its embedding-based measure of semantic diversity was 19% above the strongest competing API. These are separate measures: quality judges rated examples, while the diversity score estimates how much their meanings vary.
The comparison used different generation setups. Invent returned each dataset in one API call; for other systems the researchers chose the largest batch each could produce under a 128,000-token output limit. The report does not compare costs, so the results do not show which approach is cheaper to run. Its authors also note that the terms for some proprietary APIs prohibit using their outputs for training, even though those systems appear in the fine-tuning benchmark.
For that benchmark, the team fine-tuned three models—Llama-3.3-70B, Gemma-4-31B-it and Qwen3.5-9B—on separate 20,000-example medical question-and-answer datasets from each generator. They kept the training setup the same and asked Google’s Gemini 3.1 Pro to rank six answers to each held-out question. The Invent-data version ranked first for 54% of Llama questions, against 25% for the untuned base model; it led the Gemma comparison at 41%. For Qwen, the untuned base ranked first more often than the Invent-data version, 44% to 38%.
The study scores quality and relevance separately; relevance asks whether examples match the requested task, subject, language and stated constraints. At 5,000 examples, Invent scored 7.47 out of 10 on relevance, below Claude Opus 5 at 8.86 and GLM-5.3 at 8.66. Separately, Gemini 3.1 Pro’s medical-answer rankings are not measured medical accuracy or human preference. The report is company-authored and has no external replication, so it supports a narrower result: Invent data ranked first for two of three tested model families, but not the third.
Illustrative photo by Rahul Mishra via Unsplash, used under the Unsplash License. The image depicts a generic coding workspace, not the study or its results.