How High‑Quality Inference Data Is Powering the Next AI Revolution

In the rapid wave of AI development, data is the core driver, and the paradigm shift toward high‑quality inference data is pushing technological breakthroughs. Ant Group's Ant Financial Science (蚂蚁数科) has conducted deep research and practice in this field, covering inference data paradigms, financial scenario applications, intelligent labeling, quality inspection, and the AI Data Synthesis and Production Platform (AIGD).

1. High‑Quality Inference Data Becomes a New Industry Paradigm

The evolution of AI can be divided into three stages. Before 2022, decision‑oriented AI dominated, focusing on large‑scale descriptive data to optimize decisions. At the end of 2022, generative AI (e.g., OpenAI’s GPT‑3) sparked the era of knowledge injection, making data annotation the key to model capability. Looking ahead, AGI will require data synthesis, as traditional internet‑sourced data can no longer meet the growing demand for volume and quality.

Since the 2025 Spring release of DeepSeek R1, inference models have exploded, demanding small‑scale, high‑quality inference data with long chain‑of‑thought (CoT) reasoning. Compared with the massive labeled corpora used in generative AI, inference models rely on concise, reasoning‑rich datasets to boost logical performance.

2. Inference Data in Financial Scenarios

Financial data can be split into two categories: (1) strong‑logic datasets for calculations and logical answers, requiring deep analytical ability; (2) weaker‑logic datasets for dialogues and knowledge Q&A. Existing QA‑style datasets lack explicit reasoning steps. Ant Financial Science hypothesizes that converting them into long CoT data will significantly improve model performance.

To this end, a production pipeline for financial long‑CoT data was built, consisting of data synthesis and evaluation. The synthesis uses two steps: “from result to cause” (leveraging counterfactual reasoning) and “from cause to result” (using large models to expand data). Quality checks include causal consistency, fluency, and expert‑driven ranking and rewriting, ensuring high‑quality long‑CoT data for downstream training.

3. AIGD: AI Data Synthesis and Production Platform

The self‑developed AIGD platform creates a complete data product lifecycle covering data collection, processing, labeling, synthesis, and quality assessment. It supplies diverse data for large‑model lifecycles: massive general data for pre‑training, expert‑labeled vertical data for fine‑tuning, and long‑CoT data for inference‑model post‑training.

AIGD also supports AI safety and intelligent‑agent applications. Recent releases include the financial inference model Agentar‑Fin‑R1, the CoT dataset Agentar‑DeepFinance‑100K, and the Agentar development platform, accelerating large‑model deployment in finance.

4. Q&A

Q1: How does the reverse‑rewriting method ensure correct element extraction, causal graph alignment, and rewrite accuracy? A1: Quality control uses multi‑model voting and expert annotation to guarantee causal plausibility and sentence fluency.

Q2: How is the data‑filtering pipeline’s operator system iteratively updated? A2: AIGD provides an operator marketplace where generic and vertical operators are stored; developers assemble pipelines freely and add new operators when needed.

Q3: How is high‑quality data validated? Which metrics are used during labeling? A3: Industry experts rank multiple reasoning paths for the same question, assess reasonableness and quality, and rewrite if necessary, using expert judgment as the quality ceiling.

Overall, high‑quality inference data, especially long‑CoT data, is becoming a decisive factor for advancing AI reasoning capabilities across domains such as finance.