In the industrial pre-training of large language models (LLMs), large-scale web content extraction faces a critical trade-off: rule-based methods sacrifice quality for efficiency, while model-based methods do the opposite. As a result, existing solutions deployed at scale often lead to significant data quality degradation and corpus loss. To address this challenge, which we faced directly at Ant Group, this paper introduces HiParse, a hierarchical adaptive web parsing framework deployed in production. The core of HiParse is to abstract this multi-dimensional challenge into a ''PESO'' (Precision, Efficiency, Scope, Overhead) evaluation model, quantifying parser capabilities and parsing requirements from a ''supply-demand'' perspective. The framework integrates a hybrid parser system and uses a lightweight decision engine to dynamically route different web pages to the optimal parsing strategy. Comprehensive experiments show that HiParse not only achieves superior parsing quality (9.03/10 score) but its direct output corpus, with only 30 billion tokens, also improved the average performance of a downstream LLM on 11 benchmarks by 7.6%. Furthermore, the framework has proven its high throughput, cost-effectiveness, and alignment with business parsing demands under real industrial loads. HiParse provides a high-fidelity, scalable, and cost-effective data solution for enterprise-level LLM pre-training. All the relevant models in the paper will be open source soon.

Focusing on the hoisting mechanism of crane hydraulic, this paper gives the fault mechanism analysis and fault probability calculation method based on the probabilistic SDG model. The fault mechanism is described by the probabilistic SDG model while the fault propagation is presented by the conditional probability. And so the connection tree algorithm and figure elimination algorithm can be used in Bayesian inference to calculate the fault probability. In case of the given fault, the fault probabilities of the part components can be given.

Since 1990s, a large amount of lane detection systems have been designed for comparatively simple road condition on highway. In this paper, we propose a real-time lane detection algorithm in some complex conditions, including colored lane marks and roads with special traffic marks. The algorithm presented in this paper is based on a hyperbola model, which was proposed in our group's previous work. Tracking information is used to avoid the negative influence of arrow marks, and the colored lane marks are detected after image enhancement. The algorithm has been tested, and the experimental results demonstrate its high performance in these complex conditions

Trie hashing (TH), a primary key access method for storing and accessing records of dynamic files, is discussed. The key address is computed through a trie. A key search usually requires only one disk access when the trie is in core and two disk accesses for very large files when the trie must be on disk. A refinement to trie hashing, trie hashing with controlled load (THCL), is presented. It is designed to control the load factor of a TH file as tightly as that of a B-tree file, allows high load factor of up to 100% for ordered insertions, and increases the load factor for random insertions from 70% to over 85%. It is shown that these properties make trie hashing preferable to a B-tree.>