Data Engineering, Automated Evaluation, and Knowledge Graph Integration in Large Model Development

1. Data Engineering for Large Models The article introduces data‑centric AI, describing training data pipelines, inference data development, and data maintenance, and outlines key questions such as data requirements, sources, processing, evaluation, and versioning.

2. Model Landscape and Pre‑training Data It reviews the evolution of GPT‑1/2/3/4 and other major models, highlighting parameter sizes, data composition (Common Crawl, WebText, books, Wikipedia, etc.), multilingual capabilities, and differences between English and Chinese pre‑training corpora.

3. Automated Model Evaluation Three evaluation approaches are covered: manual business‑oriented scoring, downstream task benchmarks, and using LLMs (e.g., ChatGPT) for scoring, as well as crowdsourced voting systems like Elo‑based rankings.

4. Knowledge Graph and LLM Integration The article compares knowledge graphs and LLMs, discusses how LLMs can assist in KG construction (schema generation, data labeling, extraction, reasoning, QA) and how KGs can enhance LLM training, inference, and post‑processing through embedding injection, retrieval‑augmented generation, and external tool integration.

5. Practical Workflow and Quality Control It outlines a data‑centric workflow including site filtering, privacy and sensitivity filtering, deduplication, topic modeling, quality scoring, and version control, and presents methods for data sampling, annotation standards, and evaluation metrics.

6. Conclusions The article emphasizes that large‑model progress relies on high‑quality, large‑scale, diverse data, that domain‑specific fine‑tuning is inevitable, and that knowledge graphs should find a symbiotic role alongside LLMs in future AI systems.