90 / 90

第 90 天：作品集組裝、出處稽核與後續學習計畫

為已驗收的學習專案組裝可複現的證據作品集，重跑公開驗證命令，並以誠實、循證的計畫為課程收官。

自主學習與 Python 480 分鐘有效研習（8 小時）
01 120 分鐘

作品集清單：出處、可複現性、主張與局限

Agent 輔助工程檢查

前置要求

第 89 天掌握台帳；已完成的學習專案鏈。

學習目標

定義每個作品集條目必須滿足的清單——主張、證據路徑、複現命令、觀測結果、成本/授權狀態、局限——以已驗收專案證據和業界出處格式為據。

操作指引

1. 條目必備欄位：{project, claim, evidence_path, verify_command, observed_result, status, limitations}。status 詞表固定：verified-local（本機命令已驗證）、verified-disposable（一次性環境內已驗證）、blocked-awaiting-authorization（需要尚未授予的批准/付費/即時存取）。
2. 已驗收證據到條目對映：
- Autograd 專案（工單 17）：projects/autograd engine + gradcheck；經 python3 projects/autograd/cli.py forward-backward 與 gradcheck 為 verified-local。
- 可續訓預訓練（工單 18/19）：projects/pretrain/artifacts/run_artifact.json——10 步 CPU smoke，train_loss 4.0316，val_loss 4.0372，7440 參數；verified-local。雲 GPU 預訓練證據：blocked-awaiting-authorization。
- 真LoRA 機制（工單 20/21）：projects/lora/artifacts——loss 4.0976→4.0945、perplexity 60.20→60.01、merge_verified true（僅機制驗證）；雲前後對比：依 comparison_contract.json 為 blocked-awaiting-authorization。
- 可評測 Agent（工單 22/23）：projects/agent_learning deterministic-v1 評分器+夾具；verified-local。真實模型評測：依 live_eval_contract.json 為 not-run-awaiting-authorization。
- CI 交付練習（工單 24）：scripts/ci_practice_runner.py + tests/engineering/test_ci_practice.py；verified-disposable。
3. 誠實規則（逐字）：不得把受阻項升級為已驗證；每條主張須引用提交雜湊與工件路徑；smoke 規模結果不得當作收斂呈現。
4. 模型卡與資料說明書約定了業界期待的局限欄位——借鑑其欄位，而非抄其文本。

自主實作與記錄

起草 checklist.md：列每條目必備 6 欄位、3 個 status 值、2 條誠實規則，並標註每個已驗收專案對應的 status。

獨立掌握標準

清單覆蓋全部 6 欄位、3 種狀態，並正確標註每個已驗收專案的已驗證與受阻待授權證據。

權威來源資源與權利證明

linking CC BY 4.0

The Turing Way: A Handbook for Reproducible, Ethical and Collaborative Data Science

作者 / 機構: The Turing Way Community | 原始出處連結 ↗

對應章節 / 時間戳: Reproducible Research chapter: version control, testing, environment capture

驗證版本與日期: Zenodo DOI 10.5281/zenodo.3233853, continuously updated (Verified: 2026-10-30)

入選理由:

Community-maintained CC-BY handbook codifying reproducibility checklists (versioned code, environment capture, rerunnable commands); directly defines the Day 90 reproduction-verification and provenance-audit workflow.

來源前置條件:

Basic version control and command-line familiarity.

安全回退方案（離線彈性）:

The portfolio checklist and verification log schema are authored locally; reproduction commands run without the handbook.

對比候選資源與選型分析 (2)
Wilson et al. Good enough practices in scientific computing (PLOS CB 2017) ↗ compared_alternative [CC BY 4.0]

Concise practices paper; The Turing Way's maintained, checklist-driven coverage of environments and reproducibility is broader and kept current.

Sandve et al. Ten Simple Rules for Reproducible Computational Research (PLOS CB 2013) ↗ compared_alternative [CC BY 4.0]

Foundational rules list; The Turing Way expands the same rules into actionable checklists used verbatim for the Day 90 verification run.

linking arXiv non-exclusive license / Open Access

Model Cards for Model Reporting

作者 / 機構: Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, Timnit Gebru | 原始出處連結 ↗

對應章節 / 時間戳: Section 4 Model Card Sections (4.1-4.9)

驗證版本與日期: FAT* 2019 / arXiv:1810.03993v2 (Verified: 2026-10-30)

入選理由:

Primary standard for structured, transparent model reporting, mandating explicit disclosure of intended uses, evaluation slices, out-of-scope tasks, ethical considerations, and negative results.

來源前置條件:

Evaluation reporting, demographic and performance slicing, model limitations and caveats.

安全回退方案（離線彈性）:

Complete Model Card Markdown templates, section schemas, and evaluation disclosure checklists are provided in the curriculum.

對比候選資源與選型分析 (2)
Bommasani et al. The Foundation Model Transparency Index ↗ compared_alternative [Open Access]

Corporate transparency audit index; Mitchell et al. provides the concrete technical artifact specification for individual model releases.

Pushkarna et al. Data Cards: Purposeful and Transparent Dataset Documentation for Responsible AI ↗ compared_alternative [Open Access]

Dataset documentation focus; Mitchell et al. provides the required specification for trained model artifact reporting.

02 120 分鐘

組裝作品集清單

獨立掌握檢查（無 AI 輔助）

前置要求

第 90 天活動 1 清單；輔助條件：獨立。

學習目標

撰寫 portfolio_manifest.json，把每條學習主張對映到儲存庫工件路徑、複現命令、誠實 status 與局限說明——無輔助完成，每條路徑在儲存庫內可解析。

操作指引

1. 條目模式：{project, claim, evidence_path, verify_command, observed_result, status, limitations}。
2. 每個已驗收專案（autograd、pretrain、lora、agent_learning、ci-practice）各寫一條，另加一條課程本身。
3. 每個 evidence_path 必須能在儲存庫中解析；記錄前用 ls 或 test -f 驗證。
4. 每條主張引用觀測值：如 autograd gradcheck 的 max-abs-diff 在容差內；pretrain run_artifact.json 的 total_parameters=7440、val_loss=4.0372；lora eval_results.json 的 loss_delta=0.0031、merge_verified=true；agent_learning 對 summarize-notes 夾具的評分輸出；ci_practice_runner 的階段證據。
5. limitations 欄位必填且具體：如「10 步 CPU smoke 只驗證管線而非收斂」「確定性配接器不是真實模型」「合成指令資料不是基準」「隨機權重 smoke 機制不是品質證據」「契約校驗不是已授權雲端執行」。

自主實作與記錄

輸出至少含 5 個專案條目+1 個課程條目的 portfolio_manifest.json；斷言每個 evidence_path 存在且 status 使用固定詞表。

獨立掌握標準

清單解析到真實工件路徑、誠實狀態與每條主張的明確局限——無輔助產出。

權威來源資源與權利證明

linking arXiv non-exclusive license / Open Access

Datasheets for Datasets

作者 / 機構: Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, Kate Crawford | 原始出處連結 ↗

對應章節 / 時間戳: Section 3 Questions and Workflow (3.1 Motivation through 3.7 Maintenance)

驗證版本與日期: Communications of the ACM 2021 / arXiv:1803.09010v8 (Verified: 2026-10-30)

入選理由:

Foundational literature introducing standardized documentation for machine learning datasets, specifying questions on dataset composition, collection mechanisms, preprocessing, and distribution rights.

來源前置條件:

Dataset provenance, curation motivation, preprocessing transparency, maintenance lifecycle.

安全回退方案（離線彈性）:

Datasheet questionnaires, provenance schemas, and dataset auditing checklists are completely integrated into lesson materials.

對比候選資源與選型分析 (2)
Bender & Friedman. Data Statements for Natural Language Processing: Toward Mitigating System Bias and Enabling Better Science ↗ compared_alternative [CC BY 4.0 (ACL Anthology, 2016 or later)]

Linguistic demographic documentation focus; Gebru et al. provides the universal engineering datasheet framework adopted across AI.

Holland et al. The Dataset Nutrition Label: A Framework to Drive Higher Data Quality Standards ↗ compared_alternative [Open Access]

Nutrition label metaphor; Gebru et al. is a widely cited framework for dataset documentation in both academic and industrial practice.

linking arXiv non-exclusive license / Open Access

Model Cards for Model Reporting

作者 / 機構: Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, Timnit Gebru | 原始出處連結 ↗

對應章節 / 時間戳: Section 4 Model Card Sections (4.1-4.9)

驗證版本與日期: FAT* 2019 / arXiv:1810.03993v2 (Verified: 2026-10-30)

入選理由:

Primary standard for structured, transparent model reporting, mandating explicit disclosure of intended uses, evaluation slices, out-of-scope tasks, ethical considerations, and negative results.

來源前置條件:

Evaluation reporting, demographic and performance slicing, model limitations and caveats.

安全回退方案（離線彈性）:

Complete Model Card Markdown templates, section schemas, and evaluation disclosure checklists are provided in the curriculum.

對比候選資源與選型分析 (2)
Bommasani et al. The Foundation Model Transparency Index ↗ compared_alternative [Open Access]

Corporate transparency audit index; Mitchell et al. provides the concrete technical artifact specification for individual model releases.

Pushkarna et al. Data Cards: Purposeful and Transparent Dataset Documentation for Responsible AI ↗ compared_alternative [Open Access]

Dataset documentation focus; Mitchell et al. provides the required specification for trained model artifact reporting.

03 120 分鐘

可複現性驗證：重跑公開專案命令

Agent 輔助工程檢查

前置要求

第 90 天活動 2 清單；輔助條件：agent_assisted（Agent 可協助捕獲日誌，但驗證證據是命令輸出本身）。

學習目標

在乾淨 shell 中重跑每個專案的公開驗證命令、捕獲輸出，確認作品集主張可複現——如實記錄環境細節與任何漂移。

操作指引

1. 先記錄環境：python3 --version、uname -a（或等價物）、git rev-parse --short HEAD。
2. 執行各公開命令並捕獲 stdout+退出碼：
- python3 projects/autograd/cli.py forward-backward
- python3 projects/autograd/cli.py gradcheck
- python3 -m unittest tests.unit.test_autograd_engine
- python3 -m projects.agent_learning.cli evaluate --fixture summarize-notes
- python3 projects/lora/evidence/validate_contract.py（--require-authorized 在批准缺失時*應當*失敗——把拒絕記為正確行為）
- python3 scripts/audit_calendar_ledger.py
3. 把每個輸出與已記錄工件欄位對比（gradcheck 容差、台帳合計、契約 status blocked-awaiting-authorization）。每條命令跑兩次——一次冷啟動、一次讀過工件後——對輸出做 diff。
4. 記錄 verification_log.json：{command, exit_code, key_output, artifact_field_match: true|false, notes}。任何漂移（路徑變更、版本差異、工件缺失）須記錄不得隱藏。

自主實作與記錄

輸出覆蓋全部 6 組命令的 verification_log.json；每列帶實際觀測的輸出摘要與匹配判定。

獨立掌握標準

全部複現命令已執行並捕獲輸出；預期拒絕正確標記；不匹配如實記錄不作隱藏。

權威來源資源與權利證明

linking CC BY 4.0

The Turing Way: A Handbook for Reproducible, Ethical and Collaborative Data Science

作者 / 機構: The Turing Way Community | 原始出處連結 ↗

對應章節 / 時間戳: Reproducible Research chapter: version control, testing, environment capture

驗證版本與日期: Zenodo DOI 10.5281/zenodo.3233853, continuously updated (Verified: 2026-10-30)

入選理由:

Community-maintained CC-BY handbook codifying reproducibility checklists (versioned code, environment capture, rerunnable commands); directly defines the Day 90 reproduction-verification and provenance-audit workflow.

來源前置條件:

Basic version control and command-line familiarity.

安全回退方案（離線彈性）:

The portfolio checklist and verification log schema are authored locally; reproduction commands run without the handbook.

對比候選資源與選型分析 (2)
Wilson et al. Good enough practices in scientific computing (PLOS CB 2017) ↗ compared_alternative [CC BY 4.0]

Concise practices paper; The Turing Way's maintained, checklist-driven coverage of environments and reproducibility is broader and kept current.

Sandve et al. Ten Simple Rules for Reproducible Computational Research (PLOS CB 2013) ↗ compared_alternative [CC BY 4.0]

Foundational rules list; The Turing Way expands the same rules into actionable checklists used verbatim for the Day 90 verification run.

linking BSD-3-Clause

PyTorch Tutorial: Saving and Loading Models

作者 / 機構: PyTorch Documentation Team | 原始出處連結 ↗

對應章節 / 時間戳: Saving & Loading a General Checkpoint for Inference and/or Resuming Training

驗證版本與日期: PyTorch 2.x Documentation (Verified: 2026-10-30)

入選理由:

Official documentation detailing state_dict serialization, optimizer state saving, and safe loading patterns.

來源前置條件:

Python, PyTorch tensors and nn.Module basics.

安全回退方案（離線彈性）:

All checkpoint serialization semantics, dictionary schema keys, and safe recovery steps are self-contained in lesson code.

對比候選資源與選型分析 (2)
TensorFlow tf.train.Checkpoint Guide ↗ compared_alternative [Apache 2.0]

TensorFlow trackable checkpoint architecture; PyTorch's explicit dictionary is our target implementation standard.

Hugging Face Safetensors Documentation ↗ compared_alternative [Apache 2.0]

Safetensors provides safe zero-copy tensor storage without pickle, compared directly when discussing pickle execution risks.

04 120 分鐘

終審複習與循證後續學習計畫

獨立掌握檢查（無 AI 輔助）

前置要求

第 90 天活動 1-3 完成；第 89 天掌握台帳。

學習目標

產出最終課程評審與後續學習計畫：錨定台帳剩餘缺口的間隔複習排期、接入保留的進階年度路線、以及誠實的結課聲明。

操作指引

1. 總結掌握台帳：各域前後小計、帶恢復路徑的剩餘缺口。
2. 為每個仍未達項安排 +1/+3/+7/+30 天的間隔檢索——每次給具名資源與新變體任務（不是同一題）。
3. 把後續學習對映到保留的進階年度路線（導覽連結：歷史年度路線 / legacy roadmap at /lessons/index.html）：擴展機率統計、更大規模預訓練、部署與服務——90 天範圍外主題。
4. 按其本意寫下結課聲明：完成 90 天課程產出的是已演示任務的學習證據——不是官方認證、不保證人人三個月掌握、輔助結果不能替代獨立能力。未達標準保持可見未達。
5. 輸出 next_learning_plan.json：{remaining_gaps[], scheduled_reviews[], annual_route_links[], honest_scope_statement}。

自主實作與記錄

產出 next_learning_plan.json：為每個剩餘缺口排好複習計畫，並逐字記錄誠實範圍聲明。

獨立掌握標準

終審文件、間隔排期、年度路線連結與誠實範圍聲明完成——未達標準保持可見未達。

權威來源資源與權利證明

linking Journal publisher copyright (all rights reserved) - lawful external linking only

Improving Students' Learning With Effective Learning Techniques: Promising Directions From Cognitive and Educational Psychology

作者 / 機構: John Dunlosky, Katherine A. Rawson, Elizabeth J. Marsh, Mitchell J. Nathan, Daniel T. Willingham | 原始出處連結 ↗

對應章節 / 時間戳: Psychological Science in the Public Interest 14(1): 4-58; practice testing and distributed practice sections

驗證版本與日期: Psychological Science in the Public Interest, 2013 (Verified: 2026-10-30)

入選理由:

Detailed comparative review rating practice testing and distributed practice as high-utility techniques while rating highlighting and rereading low-utility; anchors the retrieval, correction, and spaced-review design of Days 88-90.

來源前置條件:

None beyond general academic reading.

安全回退方案（離線彈性）:

Curriculum restates the utility ratings and application conditions in original prose; every task remains runnable offline.

對比候選資源與選型分析 (2)
Pashler et al. Learning Styles: Concepts and Evidence ↗ compared_alternative [Publisher copyright (originally Wiley; now SAGE Publications)]

Important negative result on learning-styles matching but narrower scope; Dunlosky et al. rates a broader set of techniques directly usable for this curriculum.

Brown, Roediger & McDaniel: Make It Stick: The Science of Successful Learning ↗ compared_alternative [Publisher copyright (Harvard University Press)]

Strong trade-book synthesis of the same evidence base; the peer-reviewed review is preferred for citation precision and technique-by-technique utility ratings.

linking CC BY-SA 4.0 (book text) / modified MIT (sample code) share-alike

Dive into Deep Learning (d2l.ai)

作者 / 機構: Aston Zhang, Zachary C. Lipton, Mu Li, Alexander J. Smola | 原始出處連結 ↗

對應章節 / 時間戳: Section: Model Selection, Underfitting, and Overfitting; Chapter: Optimization Algorithms

驗證版本與日期: Current online edition (Verified: 2026-10-30)

入選理由:

Primary open reference tying evaluation discipline (model selection, underfitting, overfitting) and optimization algorithms (gradient descent, minibatch) to the rest of the curriculum.

來源前置條件:

Linear algebra, calculus, and probability at the Days 19-28 level.

安全回退方案（離線彈性）:

All required formulas and worked numbers are reproduced as original explanations; the book is further reading.

對比候選資源與選型分析 (2)
Deep Learning (Goodfellow, Bengio, Courville) ↗ compared_alternative [MIT Press; free HTML for reading, all rights reserved]

Detailed treatment of optimization difficulties (chapter 8) including saddle points and plateaus; all-rights-reserved text kept as a linking-only alternative.

fast.ai: Practical Deep Learning for Coders ↗ compared_alternative [Course site copyright / Apache-2.0 (fastai code)]

Strong practitioner framing, but organized around the fastai library and video lessons rather than self-contained mathematics.