Starred repositories
GSM8K evaluation: Qwen-0.6B/1.7B/4B × 3 schemes (Baseline, KV-Reuse, M-Matrix)
GPU k-NN search and k-NN graph construction on Apple silicon, built on MLX
cuVS - a library for vector search and clustering on the GPU
本项目面向中文电商图文检索这一真实问题，以 Chinese-CLIP ViT-B/16 为基础模型，在 MUGE 电商检索数据集（官方验证集）上系统比较了四类轻量化适配方法：零样本直接推理、线性探针、轻量 Adapter 与 LoRA 部分参数微调，并额外考察系统性提示模板优化。全部实验在单张 8GB 显存的 RTX 5060 Laptop GPU 上完成。
DeepSeek 4 Flash and PRO local inference engine for Metal, CUDA and ROCm
《大模型推理技术详解》— 基于 ds4 (DwarfStar) 源码的大模型推理技术教程，26 章 + 4 附录
LLM驱动的 A/H/美股智能分析：多数据源行情 + 实时新闻 + LLM决策仪表盘 + 多渠道推送，零成本定时运行，纯白嫖. LLM-powered stock analysis system for A/H/US markets.
Zhihaoi / Agent-libOS
Forked from yingqi-z20/Agent-libOSAgent libOS provides a runtime substrate for self-evolving LLM agents with process-like execution, tool management, and security boundaries implemented through explicit capabilities and runtime pri…
Official code for "MobileForge: Annotation-Free Adaptation for Mobile GUI Agents with Hierarchical Feedback-Guided Policy Optimization"
[NeurIPS'25] Listwise Preference Diffusion Optimization for User Behavior Trajectories Prediction
Toolkit for generative retrieval on Korean law DBs with constrained decoding and reproducible evaluation.
Does exposure bias in logged feedback change offline conclusions about recommender models? Retrieval, ranking and generative recommendation on KuaiRand / KuaiRec, checked against random-exposure logs.
Semantic-ID generative retrieval (TIGER) and an HSTU-style ranker in a two-stage cascade on Amazon 2023 Electronics, with a serving benchmark on one RTX 4090
Unified Schema-Based Information Extraction
Turn any LLM into a Jev-style decision model: typed decisions, real probabilities, no training. (continue updating, welcome any issue and PR request)