Source-linked AI summary
Scaling Synthetic Data Creation with 1,000,000,000 Personas
Tao Ge, Xin Chan, Xiaoyang Wang, Dian Yu, Haitao Mi, Dong Yu
TL;DR
Existing synthesis methods struggle to produce diverse data at scale because they rely on limited seed corpora or manually curated key points. The paper introduces Persona Hub, a billion-persona collection used to steer LLMs across multiple data-generation scenarios, and reports strong mathematical-reasoning results while identifying security and capability limitations.
Problem
Creating large quantities of synthetic data is straightforward, but ensuring that diversity scales requires more diverse prompts than existing seed-corpus and key-point approaches can practically provide.
Method
The paper derives one billion personas from web data and interpersonal relationships, then integrates them into data-synthesis prompts to steer LLMs toward varied perspectives.
Results
Persona Hub supports synthetic creation across mathematical and logical reasoning, instructions, knowledge-rich texts, game NPCs, and tools; a model trained on 1.07M synthetic math problems achieved nearly 80% accuracy in an in-distribution evaluation and 64.9% accuracy on MATH.
Takeaways & Limitations
Persona-driven synthesis is presented as versatile, scalable, flexible, and easy to use across varied synthetic-data scenarios.
Takeaways & Limitations
The approach could enable extraction and replication of a target LLM’s knowledge, intelligence, and capabilities when instructions broadly cover its use cases.
Abstract
from arXiv · showhide
We propose a novel persona-driven data synthesis methodology that leverages various perspectives within a large language model (LLM) to create diverse synthetic data. To fully exploit this methodology at scale, we introduce Persona Hub -- a collection of 1 billion diverse personas automatically curated from web data. These 1 billion personas (~13% of the world's total population), acting as distributed carriers of world knowledge, can tap into almost every perspective encapsulated within the LLM, thereby facilitating the creation of diverse synthetic data at scale for various scenarios. By showcasing Persona Hub's use cases in synthesizing high-quality mathematical and logical reasoning problems, instructions (i.e., user prompts), knowledge-rich texts, game NPCs and tools (functions) at scale, we demonstrate persona-driven data synthesis is versatile, scalable, flexible, and easy to use, potentially driving a paradigm shift in synthetic data creation and applications in practice, which may have a profound impact on LLM research and development.
1 Introduction
Scaling synthetic data requires many diverse prompts, but seed corpora and curated key-point lists do not practically scale. The paper proposes persona-driven synthesis and Persona Hub, a billion-persona collection for diverse data creation across scenarios.
- Synthetic data can be generated by LLMs from a specified synthesis prompt, but scaling quantity does not ensure comparable diversity.
- Prior instance-driven methods mainly inherit diversity from limited seed corpora, making it difficult to extend and scale beyond those seeds.
- Key-point-driven methods require comprehensive lists that are prohibitively difficult to enumerate across granularities outside narrow domains.
- Persona-driven synthesis steers an LLM toward distinctive perspectives, enabling broad synthetic-data coverage when paired with a comprehensive persona collection.
- 1 billion diverse personas were automatically curated into Persona Hub, representing approximately 13% of the world’s population and serving as distributed carriers of world knowledge.
- Persona Hub supports large-scale creation of mathematical and logical reasoning problems, instructions, knowledge-rich texts, game NPCs, and tools.
2 Persona Hub
Persona Hub is built by deriving personas from web text and interpersonal relationships, then deduplicating and filtering them. This process yields over one billion personas with adjustable granularity and expanded coverage.
- Text-to-Persona: Text-to-Persona infers likely readers, writers, or others associated with web text, using detailed inputs to produce fine-grained persona descriptions.
- Text-to-Persona: Persona descriptions can vary from coarse roles such as computer scientist to specific professional profiles focused on defined knowledge and skills.
- Persona-to-Persona: Persona-to-Persona supplements web-derived personas by using interpersonal relationships to derive less visible personas, including children, beggars, and behind-the-scenes crew members.
- Persona-to-Persona: Six iterations of relationship expansion enrich the collection from personas obtained through Text-to-Persona.
- Deduplication: Deduplication thresholds can be tightened when fewer instances but higher diversity are required, such as discarding personas with similarity greater than 0.5.
- Deduplication: 1,015,863,523 personas remained after MinHash and embedding-based deduplication followed by heuristic filtering of low-quality descriptions.
3 Persona-driven Synthetic Data Creation
Persona-driven synthesis inserts personas into prompts to steer an LLM toward different perspectives, while supporting zero-shot, few-shot, and persona-enhanced few-shot formats.
- Integrating a persona into a synthesis prompt can influence the LLM to adopt that persona’s perspective and create diverse data at billion scale.
- Zero-shot prompting uses no demonstrations and therefore fully exploits model creativity without constraints from specific examples.
- Few-shot prompting supplies demonstrations to better ensure that synthesized data meets specified requirements.
- Persona-enhanced few-shot prompting strengthens persona-driven synthesis but requires deriving a corresponding persona for every demonstration beforehand.
4 Use Cases
Persona-driven synthesis adapts across mathematical and reasoning tasks, using personas to steer content while retaining prompt control over focus and difficulty. Scaling synthetic mathematics to 1.07M training problems produced strong MATH performance and high sampled validity, while persona similarity predicted—but did not determine—problem similarity.
- Scope: Persona-driven synthesis supports math, logical reasoning, instructions, knowledge-rich texts, game NPCs, and tool development.The methodology is presented as general and adaptable by changing the data synthesis prompt.
- Mathematical synthesis: A linguist persona steers math problems toward computational linguistics while prompts can still specify focus or difficulty.Examples include geometry-focused and Olympiad-level prompts.
- Mathematical synthesis: Mathematics professionals yield more challenging problems by invoking deeper and more fine-grained mathematical knowledge and skills.Such personas can cover advanced mathematical concepts harvested from public web texts.
- Mathematical synthesis: 1.07M synthetic math problems were created from 1.09M personas using GPT-4 zero-shot prompting, without benchmark instances during creation.A 20K subset was held out, with 11.6K retained after solution-consistency filtering for the synthetic test set.
- Evaluation: 64.9% accuracy was achieved on MATH by a 7B model fine-tuned on the synthetic problems using greedy decoding.The evaluation is out-of-distribution because MATH instances were not used for synthesis or training.
- Evaluation: 96.5% of 200 sampled challenging math problems were judged valid by two math experts.Seven problems were marked invalid because of issues such as insufficient or conflicting conditions.
- Diversity: Synthesized-problem similarity correlates with but remains lower than persona similarity, while finance-and-probability constraints increase problem similarity.With persona similarity 0.9, most problems generated by GPT-4o and GPT-35-Turbo fall between 0.6 and 0.75 similarity.
4.2 Logical Reasoning Problems
Persona-driven prompting generates diverse logical reasoning problems when the requirements are clearly specified. The examples include standard reasoning tasks and whimsical Ruozhiba-style problems that remain relevant to their personas.
- Logical reasoning: Logical reasoning problems can be synthesized with the persona-driven methodology.The paper presents typical examples created with various personas.
- Logical reasoning: Clearly described requirements let diverse personas produce reasoning problems that satisfy the task and remain relevant to those personas.This also applies to whimsical Ruozhiba-style problems.
- Resources: The released examples include 50,000 synthetic reasoning problems.
- Ruozhiba-style problems: Ruozhiba-style problems combine puns, polysemy, causal inversion, and homophones to embed logical traps in complex Chinese-language constructs.The paper describes these questions as intricate and challenging.
4.3 Instructions
Persona Hub can simulate diverse users to synthesize instructions representing user prompts. Zero-shot prompting maximizes persona-driven variety, while persona-enhanced few-shot prompting uses demonstrations to produce instructions more like those from real users.
- Instruction synthesis: Persona Hub simulates varied users and their typical requests for LLM assistance, producing diverse instructions.These instructions typically represent the first turn of a user–LLM conversation.
- Prompting methods: Zero-shot prompting generates instructions from different personas without relying on an existing instruction dataset.
- Prompting methods: Persona-enhanced few-shot prompting samples demonstrations from existing instruction data and infers their associated personas.The method is more complex but produces instructions that more closely resemble real-user instructions.
- Conversation synthesis: Diverse first-turn instructions can be extended into many simulated conversations for instruction-following and conversational training.Two personas can also be selected to role-play conversations between real people.
- Resources: The paper released 50,000 instructions synthesized through zero-shot and persona-enhanced two-shot prompting.
4.4 Knowledge-rich Texts
Persona-driven prompting creates knowledge-rich plain text by eliciting a persona’s knowledge and perspective. Scaling across 1 billion personas is described as covering nearly any topic at multiple levels of granularity.
- Text synthesis: A persona can prompt an LLM to write a Quora article using its knowledge, skills, experience, or insights.The resulting content is described as informative and knowledge-rich.
- Text synthesis: Persona-driven knowledge-rich text synthesis can support LLM pre-training and post-training.
- Scaling: Scaling the process with 1 billion personas can produce knowledge-rich texts covering almost any topic across varied levels of granularity.
- Example format: Quora is a question-and-answer website where users ask questions and provide answers across a wide range of topics.Its articles are often written by knowledgeable individuals, including experts in various fields.
- Alternative formats: Knowledge-rich text synthesis can also target educational reading material that a persona may find interesting.
4.5 Game NPCs
Persona Hub supports large-scale game NPC creation by projecting diverse personas into supplied game worlds. Examples cover World of Warcraft and Moonlight Blade.
- NPC creation: Persona Hub projects real-world personas into game characters when paired with a game's background and world-building information.The method is intended to reduce brainstorming effort during game design.
- Game context: The examples use game worldview, story background, and character information to frame NPC generation.
- World of Warcraft: Figure 15 presents NPC creation for World of Warcraft using personas from Persona Hub.
- Moonlight Blade: Figure 16 presents NPC creation for Moonlight Blade using Persona Hub.
4.6 Tool (Function) Development
Persona Hub can generate tools tailored to personas and anticipate user needs. Interface definitions can then be converted into code implementations and reused by an LLM.
- Tool generation: The proposed prompt asks for a high-level tool interface that gives a persona access to complex functionality an LLM struggles with.
- Tool generation: Persona Hub predicts tools users might need by simulating diverse real users and their possible requests.Pre-built tools can be called when similar requests arise, avoiding construction from scratch.
- Examples: Figure 17 shows examples of tools created with various personas, including functionality a cab driver may need to check traffic conditions.
- Implementation: Interface definitions can be converted into code implementations by calling an LLM, producing pre-built tools for future use.
5 Broad Impact and Ethical Concerns
The paper presents Persona Hub as a way to broaden LLM data creation, simulate real-world users, and access distributed model knowledge. It also identifies limitations and security risks, including hallucination, training-data extraction, capability replication, and harder-to-detect synthetic misinformation.
- Paradigm shift: Persona Hub expands LLMs from processing existing data toward creating new data from many perspectives.
- Reality simulation: 1 billion personas can simulate diverse user reactions and behaviors for product launches, legislation, profiling, and cold-start settings.
- Broader applications: Persona Hub can support roleplay, agent collaboration, strategic reasoning, and organized virtual societies.
- Full memory access: Persona Hub uses diverse personas to transform an LLM's parameters into synthetic textual data, treating the personas as distributed carriers of world knowledge.
- Limitations: Current Persona Hub and LLMs limit synthetic-data breadth and quality because the model cannot losslessly convert memory into data amid hallucination.
- Security and competition: Extensive extraction of a target LLM's outputs risks dumping training data and replicating its knowledge, intelligence, and capabilities.
- Competitive landscape: The methodology may shift LLM competition from data advantage toward more advanced technologies.
- Ethical concerns: Diverse persona writing styles can make synthetic texts harder to distinguish from human content, potentially worsening contamination, misinformation, and fake-news risks.
6 Conclusion and Future Work
The paper concludes that Persona Hub is a general engine for scaling synthetic-data creation across scenarios, while identifying finer-grained personas and multimodal synthesis as future directions.
- Conclusion: Persona Hub contains 1 billion diverse personas automatically curated from web data and supports synthetic-data creation across varied scenarios.
- Persona refinement: The first version emphasizes major persona aspects and lacks fine-grained details such as preferences, family backgrounds, and life experiences.
- Future work: Future work includes refining persona descriptions, exploring multimodal synthetic-data creation, and testing super personas beyond existing knowledge.