Bilingual foundations and research notes on mathematics, statistics, machine learning, NLP, LLMs, model evaluation, and applied AI systems.https://techenglish.tophttps://techenglish.top/data-scaling-for-lora-domain-adaptationhttps://techenglish.top/data-scaling-for-lora-domain-adaptationA research note on how to study the effect of domain data size in LoRA-based machine translation adaptation.Mon, 24 Aug 2026 00:00:00 GMThttps://techenglish.top/my-lora-nmt-project-research-question-method-findingshttps://techenglish.top/my-lora-nmt-project-research-question-method-findingsA research-story version of my LoRA NMT project: problem framing, data, method, evaluation, findings, limitations, and future work.Mon, 24 Aug 2026 00:00:00 GMThttps://techenglish.top/terminology-evaluation-in-domain-mthttps://techenglish.top/terminology-evaluation-in-domain-mtWhy terminology should be evaluated explicitly in domain machine translation, especially when automatic metrics can hide critical technical errors.Mon, 24 Aug 2026 00:00:00 GMThttps://techenglish.top/why-lora-fits-low-resource-domain-adaptationhttps://techenglish.top/why-lora-fits-low-resource-domain-adaptationA research-facing explanation of why LoRA is a good fit for low-resource domain machine translation: controlled adaptation, lower experimental cost, and reduced overfitting risk.Mon, 24 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/data-scaling-for-lora-domain-adaptationhttps://techenglish.top/zh/data-scaling-for-lora-domain-adaptation如何研究领域数据规模对 LoRA 机器翻译适配的影响：数据多少才够，什么时候收益递减，什么时候是数据质量问题。Mon, 24 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/my-lora-nmt-project-research-question-method-findingshttps://techenglish.top/zh/my-lora-nmt-project-research-question-method-findings用研究叙事方式解释我的 LoRA NMT 项目：问题定义、数据、方法、评估、发现、局限与未来工作。Mon, 24 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/terminology-evaluation-in-domain-mthttps://techenglish.top/zh/terminology-evaluation-in-domain-mt为什么领域机器翻译必须单独评估术语，尤其是在自动指标可能掩盖关键技术错误的时候。Mon, 24 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/why-lora-fits-low-resource-domain-adaptationhttps://techenglish.top/zh/why-lora-fits-low-resource-domain-adaptation从研究角度解释为什么 LoRA 适合低资源领域机器翻译：受控适配、更低实验成本，以及较低过拟合风险。Mon, 24 Aug 2026 00:00:00 GMThttps://techenglish.top/ablation-vs-shapley-coalition-contexthttps://techenglish.top/ablation-vs-shapley-coalition-contextAblation is simple and useful, but Shapley values ask a broader question by averaging marginal contribution across many coalition contexts.Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/association-to-causal-evidence-attributionhttps://techenglish.top/association-to-causal-evidence-attributionTraining-data attribution can provide evidence of influence, but causal claims require careful interventions, retraining, controls, and uncertainty analysis.Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/attribution-units-data-sources-to-exampleshttps://techenglish.top/attribution-units-data-sources-to-examplesWhy attribution units matter: source, group, document, example, and token-level attribution answer different research questions and support different kinds of evidence.Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/coalitions-marginal-contributions-shapley-valueshttps://techenglish.top/coalitions-marginal-contributions-shapley-valuesThe basic Shapley framework for data attribution: coalition value, marginal contribution, averaging over contexts, and why the result is more stable than one ablation.Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/data-intervention-and-attribution-validationhttps://techenglish.top/data-intervention-and-attribution-validationHow to validate data attribution through deletion, correction, reweighting, counterfactual examples, retraining, and random deletion baselines.Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/instance-level-attribution-gradient-influence-tracinhttps://techenglish.top/instance-level-attribution-gradient-influence-tracinA practical comparison of instance-level attribution methods for NLP: gradient similarity, influence functions, and TracIn, including assumptions and limitations.Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/interview-notes-training-data-attribution-thesishttps://techenglish.top/interview-notes-training-data-attribution-thesisA compact interview narrative for a training-data attribution thesis: research question, method, contribution, limitations, and PhD extensions.Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/scaling-data-attribution-monte-carlo-surrogate-hierarchicalhttps://techenglish.top/scaling-data-attribution-monte-carlo-surrogate-hierarchicalWhy exact Shapley is expensive and how scalable attribution uses sampling, surrogate models, datamodels, and group-to-document-to-example hierarchies.Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/statistical-reliability-attribution-uncertaintyhttps://techenglish.top/statistical-reliability-attribution-uncertaintyAttribution scores are estimates. This note separates estimator bias, sampling variance, training randomness, evaluation uncertainty, and bootstrap confidence intervals.Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/training-data-attribution-for-nlp-researchhttps://techenglish.top/training-data-attribution-for-nlp-researchA research-oriented guide to training-data attribution: attribution units, utility functions, Shapley values, influence methods, causality, scalability, and uncertainty.Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/utility-functions-model-behaviourshttps://techenglish.top/utility-functions-model-behavioursTraining-data attribution depends on the utility function. Quality, factuality, style, safety, fairness, and terminology may point to different influential data.Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/what-is-training-data-attributionhttps://techenglish.top/what-is-training-data-attributionA first research-oriented explanation of training-data attribution for NLP and LLMs: what it explains, how it differs from feature attribution, and why it matters for evaluation and data-centric research.Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/ablation-vs-shapley-coalition-contexthttps://techenglish.top/zh/ablation-vs-shapley-coalition-contextAblation 简单且有用，但 Shapley 通过跨多个 coalition context 平均边际贡献，回答更全面的问题。Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/association-to-causal-evidence-attributionhttps://techenglish.top/zh/association-to-causal-evidence-attributionTraining-data attribution 可以提供数据影响的证据，但因果声称需要谨慎的干预、重新训练、控制条件和不确定性分析。Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/attribution-units-data-sources-to-exampleshttps://techenglish.top/zh/attribution-units-data-sources-to-examples为什么 attribution unit 很关键：数据源、数据组、文档、样本和 token 层面的归因回答的是不同研究问题，也支持不同证据强度。Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/coalitions-marginal-contributions-shapley-valueshttps://techenglish.top/zh/coalitions-marginal-contributions-shapley-values数据归因中的 Shapley 基础：coalition value、边际贡献、跨上下文平均，以及为什么它比一次 ablation 更全面。Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/data-intervention-and-attribution-validationhttps://techenglish.top/zh/data-intervention-and-attribution-validation如何通过删除、修正、重加权、counterfactual examples、重新训练和随机删除 baseline 来验证数据归因。Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/instance-level-attribution-gradient-influence-tracinhttps://techenglish.top/zh/instance-level-attribution-gradient-influence-tracin面向 NLP 的样本级归因方法比较：gradient similarity、influence functions 与 TracIn，以及它们的假设和局限。Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/interview-notes-training-data-attribution-thesishttps://techenglish.top/zh/interview-notes-training-data-attribution-thesis面试中解释 training-data attribution 论文的叙事模板：研究问题、方法、贡献、局限和博士阶段扩展。Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/scaling-data-attribution-monte-carlo-surrogate-hierarchicalhttps://techenglish.top/zh/scaling-data-attribution-monte-carlo-surrogate-hierarchical为什么 exact Shapley 昂贵，以及可扩展 attribution 如何使用采样、surrogate models、datamodels 和 group→document→example 层级。Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/statistical-reliability-attribution-uncertaintyhttps://techenglish.top/zh/statistical-reliability-attribution-uncertaintyAttribution 分数是估计量。本文区分 estimator bias、sampling variance、training randomness、evaluation uncertainty 与 bootstrap confidence intervals。Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/training-data-attribution-for-nlp-researchhttps://techenglish.top/zh/training-data-attribution-for-nlp-research一篇面向研究和面试的 training-data attribution 笔记：归因单位、utility function、Shapley value、influence methods、因果性、可扩展性和不确定性。Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/utility-functions-model-behaviourshttps://techenglish.top/zh/utility-functions-model-behavioursTraining-data attribution 依赖 utility function。质量、事实性、风格、安全、公平性和术语表现可能指向不同的高影响数据。Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/what-is-training-data-attributionhttps://techenglish.top/zh/what-is-training-data-attribution从 NLP 与 LLM 研究角度解释 training-data attribution：它解释什么、它和 feature attribution 有什么不同，以及为什么它对模型评估和数据中心研究重要。Sun, 23 Aug 2026 00:00:00 GMThttps://techenglish.top/anova-for-comparing-multiple-modelshttps://techenglish.top/anova-for-comparing-multiple-modelsANOVA asks whether group-level variation is larger than within-group noise.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/attention-mechanismhttps://techenglish.top/attention-mechanismAttention as a learned way to decide what context matters for each token.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/bias-variance-tradeoffhttps://techenglish.top/bias-variance-tradeoffBias and variance explain why both too-simple and too-flexible models can fail.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/bias-variance-tradeoff-in-machine-learninghttps://techenglish.top/bias-variance-tradeoff-in-machine-learningA practical diagnosis map for underfitting, overfitting, model complexity, and generalization.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/bootstrap-resampling-for-model-evaluationhttps://techenglish.top/bootstrap-resampling-for-model-evaluationBootstrap resampling estimates uncertainty by repeatedly reusing the observed test set.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/confidence-intervals-one-score-is-not-enoughhttps://techenglish.top/confidence-intervals-one-score-is-not-enoughA model score without uncertainty is easy to read but easy to overtrust.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/correlation-vs-causation-in-ai-researchhttps://techenglish.top/correlation-vs-causation-in-ai-researchCorrelation is useful evidence, but causal claims require stronger design and stronger assumptions.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/cross-validation-for-model-evaluationhttps://techenglish.top/cross-validation-for-model-evaluationWhy one split is fragile, how K-fold works, and when cross-validation can mislead in NLP.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/effect-size-statistical-significance-is-not-enoughhttps://techenglish.top/effect-size-statistical-significance-is-not-enoughA statistically significant result can still be too small to matter in research or deployment.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/feature-engineering-vs-representation-learninghttps://techenglish.top/feature-engineering-vs-representation-learningHow traditional ML features connect to embeddings, neural networks, and modern NLP systems.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/fine-tuning-transformershttps://techenglish.top/fine-tuning-transformersHow pretrained language models are adapted to a task or domain with supervised data.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/grid-search-and-randomized-searchhttps://techenglish.top/grid-search-and-randomized-searchHow to tune hyperparameters without confusing search effort with scientific evidence.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/hypothesis-testing-without-ritualshttps://techenglish.top/hypothesis-testing-without-ritualsHypothesis testing is a way to discipline claims, not a ritual for producing p-values.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/linear-regression-as-a-statistical-modelhttps://techenglish.top/linear-regression-as-a-statistical-modelLinear regression is more than a line: it is a model of signal, noise, assumptions, and explanation.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/llm-evaluation-and-failure-modeshttps://techenglish.top/llm-evaluation-and-failure-modesA practical map of LLM evaluation risks: hallucination, prompt sensitivity, bias, contamination, and brittle benchmarks.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/map-estimation-and-bayesian-thinkinghttps://techenglish.top/map-estimation-and-bayesian-thinkingMAP estimation shows how prior beliefs and observed evidence combine.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/maximum-likelihood-estimation-why-models-learn-parametershttps://techenglish.top/maximum-likelihood-estimation-why-models-learn-parametersMaximum likelihood connects probability models to parameter learning.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/metrics-beyond-accuracyhttps://techenglish.top/metrics-beyond-accuracyAccuracy is easy to understand, but often wrong for imbalanced, ranked, or cost-sensitive tasks.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/model-selection-practical-guidehttps://techenglish.top/model-selection-practical-guideModel selection is the disciplined process of choosing among models without fooling yourself.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/multiple-comparisons-in-ai-experimentshttps://techenglish.top/multiple-comparisons-in-ai-experimentsWhen we try many models, prompts, seeds, and metrics, false discoveries become easier than they look.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/nlp-evaluationhttps://techenglish.top/nlp-evaluationWhy NLP evaluation needs metrics, uncertainty, human judgment, and task-specific error analysis.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/overfitting-and-regularizationhttps://techenglish.top/overfitting-and-regularizationHow models learn noise, how validation curves reveal it, and how regularization controls it.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/parameter-efficient-fine-tuninghttps://techenglish.top/parameter-efficient-fine-tuningWhy LoRA and adapters are useful when full fine-tuning is too expensive or unstable.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/pca-from-variance-to-representationhttps://techenglish.top/pca-from-variance-to-representationPCA explains how variance, projection, and representation are connected.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/pipelines-and-data-leakagehttps://techenglish.top/pipelines-and-data-leakageWhy preprocessing belongs inside the validation pipeline, not before the split.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/power-analysis-how-many-experimentshttps://techenglish.top/power-analysis-how-many-experimentsPower analysis connects sample size, effect size, and the chance of detecting a real improvement.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/probability-distributions-for-machine-learninghttps://techenglish.top/probability-distributions-for-machine-learningHow distributions become assumptions about data, labels, errors, and model behavior.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/prompting-and-instruction-followinghttps://techenglish.top/prompting-and-instruction-followingPrompting as task specification, interface design, and evaluation risk.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/random-variables-expectation-and-variancehttps://techenglish.top/random-variables-expectation-and-varianceA practical introduction to random variables, expectation, and variance for AI experiments.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/regularization-statistics-behind-l1-and-l2https://techenglish.top/regularization-statistics-behind-l1-and-l2Regularization controls model complexity by making some parameter values less plausible.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/sequence-to-sequence-modelshttps://techenglish.top/sequence-to-sequence-modelsThe encoder-decoder idea behind machine translation, summarization, and many generation tasks.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/statistical-testing-for-nlp-and-llm-evaluationhttps://techenglish.top/statistical-testing-for-nlp-and-llm-evaluationA practical map for choosing statistical tests in NLP, MT, RAG, and LLM evaluation.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/tokenization-and-subwordshttps://techenglish.top/tokenization-and-subwordsTokenization is the first modeling decision in NLP: it decides what units a model can see, how rare terms are represented, and how multilingual systems handle domain language.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/train-validation-test-splitshttps://techenglish.top/train-validation-test-splitsA practical guide to splitting data so model evaluation stays honest.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/transformers-for-nlphttps://techenglish.top/transformers-for-nlpHow Transformers combine self-attention, feed-forward layers, residuals, and positional information.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/why-statistics-matters-for-ai-researchhttps://techenglish.top/why-statistics-matters-for-ai-researchStatistics is not just a set of formulas. It is a way to reason about uncertainty, evidence, and trust in AI experiments.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/word-and-sentence-embeddingshttps://techenglish.top/word-and-sentence-embeddingsHow discrete language becomes vector space, and why sentence embeddings matter for retrieval and evaluation.Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/anova-for-comparing-multiple-modelshttps://techenglish.top/zh/anova-for-comparing-multiple-modelsANOVA 问的是：组间差异是否大于组内噪声。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/attention-mechanismhttps://techenglish.top/zh/attention-mechanism注意力机制是一种学习到的上下文选择方式：每个 token 应该看哪里。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/bias-variance-tradeoffhttps://techenglish.top/zh/bias-variance-tradeoff偏差和方差解释了为什么模型太简单或太灵活都会失败。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/bias-variance-tradeoff-in-machine-learninghttps://techenglish.top/zh/bias-variance-tradeoff-in-machine-learning用 bias–variance 诊断 underfitting、overfitting、模型复杂度和泛化问题。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/bootstrap-resampling-for-model-evaluationhttps://techenglish.top/zh/bootstrap-resampling-for-model-evaluationBootstrap 通过反复重采样已有测试集，估计模型分数和模型差异的不确定性。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/confidence-intervals-one-score-is-not-enoughhttps://techenglish.top/zh/confidence-intervals-one-score-is-not-enough没有不确定性的模型分数很好读，也最容易被过度相信。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/correlation-vs-causation-in-ai-researchhttps://techenglish.top/zh/correlation-vs-causation-in-ai-research相关性是有用证据，但因果结论需要更强的实验设计和假设。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/cross-validation-for-model-evaluationhttps://techenglish.top/zh/cross-validation-for-model-evaluation为什么一次划分很脆弱，K-fold 如何工作，以及交叉验证什么时候会在 NLP 中误导我们。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/effect-size-statistical-significance-is-not-enoughhttps://techenglish.top/zh/effect-size-statistical-significance-is-not-enough统计显著的结果，仍然可能小到没有研究或应用意义。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/feature-engineering-vs-representation-learninghttps://techenglish.top/zh/feature-engineering-vs-representation-learning传统 ML 特征如何连接到 embedding、神经网络和现代 NLP 系统。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/fine-tuning-transformershttps://techenglish.top/zh/fine-tuning-transformers如何用监督数据把预训练语言模型适配到任务或领域。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/grid-search-and-randomized-searchhttps://techenglish.top/zh/grid-search-and-randomized-search如何调超参数，同时避免把搜索努力误认为科学证据。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/hypothesis-testing-without-ritualshttps://techenglish.top/zh/hypothesis-testing-without-rituals假设检验不是为了制造 p-value，而是为了约束我们能说什么。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/linear-regression-as-a-statistical-modelhttps://techenglish.top/zh/linear-regression-as-a-statistical-model线性回归不只是一条线，而是关于信号、噪声、假设和解释的统计模型。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/llm-evaluation-and-failure-modeshttps://techenglish.top/zh/llm-evaluation-and-failure-modesLLM 评估风险图：幻觉、prompt 敏感性、偏见、污染和脆弱 benchmark。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/map-estimation-and-bayesian-thinkinghttps://techenglish.top/zh/map-estimation-and-bayesian-thinkingMAP 估计展示了先验信念如何和观察证据结合。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/maximum-likelihood-estimation-why-models-learn-parametershttps://techenglish.top/zh/maximum-likelihood-estimation-why-models-learn-parameters最大似然把概率模型和参数学习连接起来。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/metrics-beyond-accuracyhttps://techenglish.top/zh/metrics-beyond-accuracyAccuracy 容易理解，但在不平衡、排序或代价敏感任务中经常不够。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/model-selection-practical-guidehttps://techenglish.top/zh/model-selection-practical-guide模型选择是有纪律地在多个模型之间做选择，同时避免欺骗自己。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/multiple-comparisons-in-ai-experimentshttps://techenglish.top/zh/multiple-comparisons-in-ai-experiments当我们测试很多模型、prompt、seed 和指标时，假阳性会比想象中更容易出现。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/nlp-evaluationhttps://techenglish.top/zh/nlp-evaluation为什么 NLP 评估需要指标、不确定性、人类判断和任务特定错误分析。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/overfitting-and-regularizationhttps://techenglish.top/zh/overfitting-and-regularization模型为什么会学到噪声，验证曲线如何暴露它，正则化如何控制它。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/parameter-efficient-fine-tuninghttps://techenglish.top/zh/parameter-efficient-fine-tuning为什么当 full fine-tuning 太贵或不稳定时，LoRA 和 adapters 会有价值。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/pca-from-variance-to-representationhttps://techenglish.top/zh/pca-from-variance-to-representationPCA 连接了方差、投影和表示空间，是理解 embedding 与降维的重要入口。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/pipelines-and-data-leakagehttps://techenglish.top/zh/pipelines-and-data-leakage为什么预处理应该放进验证 pipeline，而不是在划分前对全数据提前处理。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/power-analysis-how-many-experimentshttps://techenglish.top/zh/power-analysis-how-many-experimentsPower analysis 连接样本量、效应大小，以及检测到真实提升的概率。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/probability-distributions-for-machine-learninghttps://techenglish.top/zh/probability-distributions-for-machine-learning概率分布不是抽象曲线，而是关于数据、标签、误差和模型行为的假设。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/prompting-and-instruction-followinghttps://techenglish.top/zh/prompting-and-instruction-followingPrompting 既是任务说明，也是接口设计和评估风险。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/random-variables-expectation-and-variancehttps://techenglish.top/zh/random-variables-expectation-and-variance从 AI 实验角度理解随机变量、期望和方差：分数为什么会波动，我们如何描述这种波动。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/regularization-statistics-behind-l1-and-l2https://techenglish.top/zh/regularization-statistics-behind-l1-and-l2正则化通过约束参数，让模型复杂度不至于吞掉泛化能力。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/sequence-to-sequence-modelshttps://techenglish.top/zh/sequence-to-sequence-models机器翻译、摘要和许多生成任务背后的 encoder-decoder 思想。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/statistical-testing-for-nlp-and-llm-evaluationhttps://techenglish.top/zh/statistical-testing-for-nlp-and-llm-evaluation一张面向 NLP、MT、RAG 和 LLM 评估的统计检验选择图。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/tokenization-and-subwordshttps://techenglish.top/zh/tokenization-and-subwordsTokenization 是 NLP 里的第一个建模决定：它决定模型能看见什么单位、稀有术语如何表示，以及多语系统如何处理领域语言。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/train-validation-test-splitshttps://techenglish.top/zh/train-validation-test-splits如何划分训练集、验证集和测试集，才能让模型评估保持诚实。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/transformers-for-nlphttps://techenglish.top/zh/transformers-for-nlpTransformers 如何组合 self-attention、feed-forward layers、residuals 和位置信息。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/why-statistics-matters-for-ai-researchhttps://techenglish.top/zh/why-statistics-matters-for-ai-research统计不是公式集合，而是帮助我们理解 AI 实验不确定性、证据强度和模型评估可信度的思维工具。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/zh/word-and-sentence-embeddingshttps://techenglish.top/zh/word-and-sentence-embeddings离散语言如何变成向量空间，以及句向量为什么对检索和评估重要。Wed, 19 Aug 2026 00:00:00 GMThttps://techenglish.top/why-low-resource-domain-mt-is-hardhttps://techenglish.top/why-low-resource-domain-mt-is-hardDomain-specific machine translation is not only a modeling problem. It is a data, terminology, evaluation, and risk problem.Sun, 16 Aug 2026 00:00:00 GMThttps://techenglish.top/about-this-blog-foundations-and-research-noteshttps://techenglish.top/about-this-blog-foundations-and-research-notesA short map of the blog: foundations, research applications, bilingual notes, and how this site complements my portfolio.Sat, 15 Aug 2026 00:00:00 GMThttps://techenglish.top/what-lora-actually-changeshttps://techenglish.top/what-lora-actually-changesA practical explanation of LoRA for NMT: what it changes, why it is useful for domain adaptation, and how it differs from full fine-tuning.Sat, 15 Aug 2026 00:00:00 GMThttps://techenglish.top/data-quality-diagnostics-for-parallel-corporahttps://techenglish.top/data-quality-diagnostics-for-parallel-corporaA practical note on diagnosing parallel corpora through alignment, duplicates, length ratio, completeness, terminology, and domain coverage.Fri, 14 Aug 2026 00:00:00 GMThttps://techenglish.top/vectors-matrices-and-embeddingshttps://techenglish.top/vectors-matrices-and-embeddingsA foundation note connecting linear algebra to embeddings, similarity, neural layers, and retrieval.Fri, 14 Aug 2026 00:00:00 GMThttps://techenglish.top/hyperparameter-optimization-grid-optuna-ashahttps://techenglish.top/hyperparameter-optimization-grid-optuna-ashaA practical research-engineering note on tuning LoRA rank, alpha, dropout, Pareto selection, ASHA pruning, and fANOVA analysis.Thu, 13 Aug 2026 00:00:00 GMThttps://techenglish.top/bleu-chrf-comet-practical-guidehttps://techenglish.top/bleu-chrf-comet-practical-guideA practical guide to BLEU, chrF, COMET, terminology metrics, and human error analysis for domain-specific MT.Wed, 12 Aug 2026 00:00:00 GMThttps://techenglish.top/lora-vs-full-finetuning-what-do-we-savehttps://techenglish.top/lora-vs-full-finetuning-what-do-we-saveA practical comparison of LoRA and full fine-tuning in low-resource domain MT: parameters, compute, stability, performance, and trade-offs.Tue, 11 Aug 2026 00:00:00 GMThttps://techenglish.top/research-engineering-for-nlp-experimentshttps://techenglish.top/research-engineering-for-nlp-experimentsA research-engineering note on repositories, configs, data splits, logs, evaluation tables, and reproducible experiment workflows.Fri, 07 Aug 2026 00:00:00 GMThttps://techenglish.top/multilingual-ai-as-a-system-problemhttps://techenglish.top/multilingual-ai-as-a-system-problemA research note framing multilingual AI beyond translation: data, tokenization, evaluation, domains, language variation, and deployment constraints.Thu, 06 Aug 2026 00:00:00 GMThttps://techenglish.top/bias-evaluation-is-more-than-promptinghttps://techenglish.top/bias-evaluation-is-more-than-promptingA responsible-AI note on controlled evaluation design, prompts, labels, metrics, human validation, and model-auditing evidence.Wed, 05 Aug 2026 00:00:00 GMThttps://techenglish.top/retrieval-is-not-just-vector-searchhttps://techenglish.top/retrieval-is-not-just-vector-searchA retrieval-systems note on sparse, dense, hybrid, reranking, graph expansion, long documents, and evidence grounding.Tue, 04 Aug 2026 00:00:00 GMThttps://techenglish.top/applied-ml-systems-from-data-to-demohttps://techenglish.top/applied-ml-systems-from-data-to-demoAn applied ML note on turning messy data into pipelines, models, evaluation reports, and usable demos.Mon, 03 Aug 2026 00:00:00 GMT