2026 07 17 llm training pipeline
注
本文系统梳理大语言模型（LLM）从预训练到推理的完整训练流程，涵盖预训练、微调、RL、RLHF 等核心阶段的技术原理与协作关系。本文属技术科普笔记，旨在帮助读者建立对大模型训练全貌的结构化认知。
调研分析：大模型训练全流程解析
Quick links: Home · Quick Start
Blog¶
This blog tracks design thinking, technical trade-offs, and engineering notes related to scientific data processing in the PorosData project.
These notes are "living" — they evolve with code and understanding, and may reflect new thinking earlier than formal documentation. They are written for readers interested in data processing pipelines, architecture evolution, and engineering practice.
Recent posts¶
2026 05 06 design insights
Note
Exploratory note on architecture and open questions around scientific data processing. Not a stable product commitment.
Design Insights
Quick links: Home · Quick Start
Model-centric pre-training and literature-scale mining impose different failure sensitivities. The pipeline therefore treats semantic stability, structural consistency, and minability as explicit design targets, implemented mainly through Processor (quality) and Designer (views and contracts). Executable detail on pipelines, tags, and acceptance tests now lives in those product docs rather than in this essay.
2026 05 06 research review
Note
Exploratory note on architecture and open questions around scientific data processing. Not a stable product commitment.
Research Review
Scientific papers encode terminology, numerals, notation, figures, and qualifiers in ways that general-purpose text tools rarely preserve. Once PDF or OCR pipelines emit machine-readable text, small formatting defects propagate into training corpora, extractors, and review tools—so preparation standards must be domain-aware, not merely “spell-checked.”
Mapping the pipeline to the docs
This documentation splits operational detail (how to run and validate) from conceptual pages (contracts, layout, governance). This note is a map, not a tutorial: it tells you where to read once you know which stage you care about.