Liam Liu

Throughout my research career, I have led or made significant contributions to the following widely recognized open-source projects: MAP-NEO: We successfully developed and open-sourced MAP-NEO, a fully open-source, State-of-the-Art (SOTA) bilingual (Chinese-English) large language model. MAP-NEO is renowned for its exceptional performance and complete transparency. We have publicly released all details, including datasets, cleaning pipelines, training code, and model weights, aiming to push the boundaries of large model research. PIN-14M: This is a pioneering multimodal pre-training dataset with over 300,000 downloads. PIN-14M, through its unique "Paired and Interleaved" format, provides unprecedented knowledge-intensive training data for multimodal large models, significantly enhancing their capabilities in complex reasoning and comprehension tasks. FineFineWeb: As a crucial data cornerstone for the AI community, FineFineWeb is a large-scale web text dataset, meticulously cleaned and processed. It boasts over 1 million monthly downloads, providing high-quality training corpora for researchers and developers worldwide. SuperGPQA: I participated in the construction of SuperGPQA, a new, highly challenging, and high-quality benchmark designed to more accurately assess the comprehensive capabilities of large models. This benchmark has been adopted by cutting-edge models such as Qwen3 and Seed Thinking v1.5, becoming a new standard for measuring the performance of top-tier models. YuE: As an open-source project benchmarking Suno, YuE is currently the leading open-source music generation model. It can generate complete songs, including vocals and accompaniment, based on lyrics. The project has gained significant attention on GitHub, amassing over 15,000 stars, and has inspired creativity among developers worldwide.