synthetic data
In plain wordsData created by computer programs or AI models rather than collected from the real world — such as practice math problems written by a model, or simulated street scenes for self-driving cars.
Examples or records produced by software — a simulator, procedural rules, a statistical model or a generative AI system — to stand in for, or add to, data gathered from people, sensors or the world, most often to train or evaluate other models. Long-standing uses include simulated environments for robotics and driving (see sim-to-real transfersim-to-real transfer, n.: Training a robot's AI in a computer simulation, where practice is fast, cheap and safe, and then making it work on the real robot despite the differences between simulation and reality. and domain randomizationdomain randomization, n.: Training a robot's AI in a simulation whose colors, lighting, weights, friction and other details are shuffled at random on every run, so that the real world holds few surprises it has not already practiced for.), rendered images with perfect labels, and privacy-preserving releases of statistical records that mimic real data without corresponding to real people. In language modeling it has become central: models generate instruction–response pairs (Self-InstructSelf-Instruct, n.: A technique in which a language model writes its own practice tasks and answers, starting from a handful of human examples; the results are filtered and used to train it to follow instructions., distillationknowledge distillation, n.: Training a smaller or cheaper 'student' AI to copy the behavior of a larger 'teacher' AI, so the student keeps much of the teacher's skill while being faster to run. from stronger models), rephrase web text into cleaner forms, write textbook-style material and coding exercises, and produce reasoning traces that are kept only if a verifier confirms the answer (rejection samplingrejection sampling, n.: Generating many candidate answers and keeping only those that pass a check — a scoring model, a test or a correct final answer — often so the AI can then be trained on its own best work., reinforcement learning with verifiable rewardsreinforcement learning with verifiable rewards, n.: Training an AI by letting it attempt problems whose answers can be checked automatically — a math result, a passing program — and rewarding it only when the check succeeds.). Its advantages are scale, control, low cost and relief from some privacy and copyright problems; its risks include amplifying the generator's errors and biases, loss of diversity and, if used carelessly over repeated generations, model collapsemodel collapse, n.: The gradual decline of AI models trained on the output of earlier AI models: rare details vanish first, then quality and variety erode, like a photocopy of a photocopy..
Half of our math fine-tuning set is synthetic, generated by a larger model and checked by a solver.
Origin
Donald Rubin proposed fully synthetic data in 1993 as a way to release census-style records without disclosing real respondents, and Roderick Little proposed partially synthetic data the same year; simulated training images go back to 1980s self-driving research. For language models the turn came with Self-Instruct (2022) and Microsoft's 'Textbooks Are All You Need' (2023).
Usage note
'Synthetic' does not mean free of real data: generators are trained on real data and can reproduce it, including personal or copyrighted material. Whether synthetic data can substitute for fresh human data at the frontier is debated (see data walldata wall, n.: The worry that AI companies will run out of new human-written text to train ever-larger models on, because the useful public internet has already been largely used up.).
Related
- model collapsemodel collapse, n.: The gradual decline of AI models trained on the output of earlier AI models: rare details vanish first, then quality and variety erode, like a photocopy of a photocopy.
- Self-InstructSelf-Instruct, n.: A technique in which a language model writes its own practice tasks and answers, starting from a handful of human examples; the results are filtered and used to train it to follow instructions.
- knowledge distillationknowledge distillation, n.: Training a smaller or cheaper 'student' AI to copy the behavior of a larger 'teacher' AI, so the student keeps much of the teacher's skill while being faster to run.
- sim-to-real transfersim-to-real transfer, n.: Training a robot's AI in a computer simulation, where practice is fast, cheap and safe, and then making it work on the real robot despite the differences between simulation and reality.
- data walldata wall, n.: The worry that AI companies will run out of new human-written text to train ever-larger models on, because the useful public internet has already been largely used up.
- rejection samplingrejection sampling, n.: Generating many candidate answers and keeping only those that pass a check — a scoring model, a test or a correct final answer — often so the AI can then be trained on its own best work.
- data augmentationdata augmentation, n.: Creating extra training examples by changing existing ones in ways that keep their meaning — a flipped, cropped or recolored photo of a cat still shows a cat — so the model learns more robustly.
Sources
- Rubin, D. B. (1993). Discussion: Statistical Disclosure Limitation. Journal of Official Statistics 9(2).
- Gunasekar, S. et al. (2023). Textbooks Are All You Need. arXiv preprint.
Mentioned in
knowledge distillationknowledge distillation, n.: Training a smaller or cheaper 'student' AI to copy the behavior of a larger 'teacher' AI, so the student keeps much of the teacher's skill while being faster to run., training datatraining data, n.: The examples an AI system learns from. Whatever patterns, skills, gaps and biases are in the training data tend to show up in the finished system., data augmentationdata augmentation, n.: Creating extra training examples by changing existing ones in ways that keep their meaning — a flipped, cropped or recolored photo of a cat still shows a cat — so the model learns more robustly., data walldata wall, n.: The worry that AI companies will run out of new human-written text to train ever-larger models on, because the useful public internet has already been largely used up., digital twindigital twin, n.: A detailed virtual copy of a real machine, factory, robot or city, kept in sync with live sensor data and used to monitor it, test changes safely and predict problems., instruction datasetinstruction dataset, n.: A collection of example requests paired with good answers, used to teach a pretrained language model to follow instructions and act as a helpful assistant., long taillong tail, n.: The huge number of rare things — uncommon words, unusual road situations, obscure facts — that each appear seldom but together make up a large share of what an AI system meets., model collapsemodel collapse, n.: The gradual decline of AI models trained on the output of earlier AI models: rare details vanish first, then quality and variety erode, like a photocopy of a photocopy., pretraining datapretraining data, n.: The enormous body of text, code and other material — mostly from the web — that a large AI model first learns from, before it is tuned to follow instructions or chat., recursive self-improvementrecursive self-improvement, n.: An AI improving its own design or training, so that each improved version is better at making the next improvement. It is the engine usually imagined behind an intelligence explosion., self-improvementself-improvement, n.: When an AI gets better by learning from material it produces itself — its own answers, critiques or practice problems — rather than only from new human-made examples., small language modelsmall language model, n.: A language model built to be compact enough to run cheaply and quickly, even on a laptop or phone, trading some knowledge and skill for speed, privacy and lower cost., SMOTESMOTE, abbr.: A technique for lopsided data that invents extra examples of the rare category by blending each rare example with one of its nearest rare neighbors, so a model sees more of it during training., AlphaGeometryAlphaGeometry, prop. n.: A Google DeepMind system from 2024 that solves olympiad geometry problems nearly as well as top human competitors, by pairing a language model that suggests constructions with a logic engine that checks every step., data flywheeldata flywheel, n.: A cycle in which a product's users generate data that improves its AI, the better AI attracts more users, and those users generate still more data., data-constrained scalingdata-constrained scaling, n.: How to keep improving AI models when the supply of fresh, high-quality training text runs out — for example by reusing the same data several times, which works well at first and then stops helping., domain randomizationdomain randomization, n.: Training a robot's AI in a simulation whose colors, lighting, weights, friction and other details are shuffled at random on every run, so that the real world holds few surprises it has not already practiced for., Self-InstructSelf-Instruct, n.: A technique in which a language model writes its own practice tasks and answers, starting from a handful of human examples; the results are filtered and used to train it to follow instructions., subliminal learningsubliminal learning, n.: The finding that an AI model can pass hidden traits — a favorite animal, or bad behavior — to another model through training data that seems to have nothing to do with them., weak supervisionweak supervision, n.: Training AI with labels that are cheap but imperfect, such as rough rules of thumb, partial tags or noisy guesses, instead of carefully hand-checked answers., world foundation modelworld foundation model, n.: A large AI model trained on vast amounts of video that can imagine how a scene will unfold, used to create realistic practice worlds for robots and self-driving cars.