Yiming Jia

Drove continuous post-training iteration for Nanbeige4-3B-Thinking (https://huggingface.co/Nanbeige/Nanbeige4-3B-Thinking-2511), owning the code capability track end-to-end. Performed RLVR training with Verl, running rapid experiment cycles on data, rewards, and training configurations to steadily improve code performance. Built and maintained the evaluation + regression framework for coding and software-engineering tasks, with primary benchmarks on LiveCodeBench v5/v6 and FullstackBench. Delivered measurable benchmark gains through systematic ablations, reproducible reporting, and iterative refinement of training and evaluation pipelines.

- Led the development of a novel multimodal data acquisition pipeline that successfully extracted high-quality instruction data from the web using carefully curated seed images. - Created one of the largest multimodal instruction datasets to date, containing 906K high-quality question-answer pairs (including 347K with images) spanning mathematics, physics, finance, chemistry, and engineering. - Implemented innovative data refinement techniques including structured content extraction, multi-path reasoning verification, and consistency filtering to ensure exceptional dataset quality. - Demonstrated the dataset's effectiveness by training MAmmoTH-VL2, which achieved state-of-the-art performance among 7B-parameter models across seven multimodal benchmarks (50.4% average accuracy). - Significantly advanced multimodal reasoning capabilities, with the trained model showing exceptional performance on mathematical tasks (MathVista: 68.1%) and complex reasoning benchmarks (MMMU-Pro: 40.7%). - Pioneered an effective approach for enhancing vision-language models through high-quality instruction data, addressing a critical bottleneck in multimodal AI development. Project fully open-sourced with dataset, model weights, and source code. Research paper completed and submitted to ICCV 2025 (under review).

CSC148 - Introduction to Computer Science Language: Python 1. Classroom Assist - Helping students with worksheet questions during lectures 2. Office Hour - Help students with general course content and homework questions 3. Midterm and final Marking

1. Analyzed and preprocessed time-series single-cell data using PCA, UMAP, Clustering, and FFT. 2. Proposed a unique way to represent genes with frequency domain data. 3. Devised and trained the whole pipeline integrated GCN+VAE model to derive gene embeddings. 4. Proved the strong biological significance of the embeddings through GO Enrichment Analysis. 5. Conducted ablation experiment to prove the significance of FFT.

1. Reproduced and trained CenterMask based on Google object detection API with the Fashionpedia dataset. 2. Optimized segmentation performance of the model in PC simulations. 3. Simplified, quantified, and transplanted my model to Sony IMX-500 chips. 4. Tested and improved the model's overall performance on Sony cameras. 5. Designed and implemented Convolutional Autoencoder to compress the size of segmentation output to meet the Wi-Fi module's bandwidth demands.

1. Contributed to GammaGL and implemented JK-net based on TensorLayerX. 2. Reproduced and optimized the experimental results of JK-net and APPNP in GammaGL. 3. Compared the different performances of graph neural networks on different backends and analyzed the reasons.