Real data and model source register

Modules 6–14 use actual external records and human annotations. No new synthetic illustration/training dataset is created. Earlier foundation fixtures remain labeled separately.

External datasets

wine

Cortez, P., et al. (2009). Wine Quality. UCI. DOI:10.24432/C56S3T. Official source · CC BY 4.0.

Reviewed archive SHA-256: 3ed56667f4b828242bd732d7d1dd7f2861e54432239d7fa63877014cbb0304d4.

Derived processing selects real red-wine chemistry, groups identical profiles, and declares a rating threshold for supervised cases. Clustering excludes the quality label.

adult

Becker, B. & Kohavi, R. (1996). Adult. UCI. DOI:10.24432/C5XW20. Official source · CC BY 4.0.

Reviewed archive SHA-256: 7537312dd56c2b98035880805ce99e68183a30ee468aa5329d6df0fbb3cc21bb.

Derived processing uses the provided benchmark files, removes exact full-row overlaps, fits five numeric inputs, and audits recorded groups. Person independence and a justified eligibility purpose are not established.

sms

Almeida, T. & Hidalgo, J. (2011). SMS Spam Collection. UCI. DOI:10.24432/C5CC84. Official source · CC BY 4.0.

Reviewed archive SHA-256: 1587ea43e58e82b14ff1f5425c88e17f8496bfcdb67a583dbff9eefaf9963ce3.

Derived processing hashes normalized text, removes duplicate/conflicting texts, selects a declared real sample, and preserves human labels. LLM cases mask contact-like strings and disclose tiny balanced subsets. Public messages can still contain personal content.

Acquisition and offline replay

Original archives, named original members, provenance.json, and ATTRIBUTION.txt remain under runs/datasets/name. New archives must match the reviewed digest; cached member hashes are verified. Source changes require review rather than bypassing the check.

Separate attributed caches are available for wine, SMS, and Adult. Extract each into its named runs/datasets directory, preserving all files. External records stay outside the original-source ZIP. Worked reports can be read without executing a dataset or model. If a download link does not open, use the offline access route.

Pretrained model

HuggingFaceTB/SmolLM2-135M-Instruct has an Apache-2.0 card/license. Reviewed revision: 12fd25f77366fa6b3b4b768ec3050bf629380bac. It runs locally on CPU with safetensors and remote code disabled. The course does not claim to reproduce or audit every pretraining source. Our fine-tuning illustrations use real human SMS labels, not generated records.

First use downloads several hundred MB; retain runs/models for replay. Base weights remain separate from the original-source archive. Consult the official card and SmolLM2 technical report. A pinned revision is reproducible, not a quality guarantee.

Use limits

Data rights and model licenses are distinct. Preserve credits and disclose transformations. Public access does not prove representative sampling, privacy, causal validity, consent for every new purpose, or operational suitability. The reports retain weak-model and format failures.

Authentic IT extension: website phishing

UCI Phishing Websites, Mohammad & McCluskey (2012), DOI 10.24432/C51W2X; UCI lists CC BY 4.0. The selected Training Dataset.arff contains 11,055 records and 30 encoded attributes. Archive SHA-256: 514f582cb44cba17251baf1d59f64cddff16efa2b644d091ce8d2a1e8d492940. Raw files remain unchanged; the case documents label recoding, ambiguous-profile exclusions and grouped splits. No live URLs are visited. Case workflow and limits.

Authored agent controls

data/samples/agent-requests.csv is original teaching material: labelled demonstration, validation and test requests. It contains no real user interactions and is not an external dataset. Module 14 returns statistics from the attributed UCI wine records; keep that evidence separate from the authored routing controls.

Original runbook retrieval fixture

data/samples/runbooks/ holds 12 original, fictional IT runbook passages and 32 authored questions for the Module 10 retrieval extension, released under CC BY 4.0 with the course content. They describe no real organization or system and are not operational guidance. Question labels were written by the course author; the corpus is small, so its scores illustrate the evaluation method, not retrieval quality on real documentation.

RT-IoT2022 completed-flow transfer

S., B. & Nagapadma, R. (2023). RT-IoT2022 [Dataset]. UCI Machine Learning Repository. DOI. Official source, CC BY 4.0.

The inspected archive contains 123,117 records and 12 labels. Reviewed archive SHA-256: bcaa24d62abbb1215be576d5cf9c02dfcb0bb7c4c2f5a00e03055afaa1ed109e. The study selects five completed-flow inputs, excludes the unnamed index and derives a binary attack target. Identical selected-input profiles remain together; capture-session, device and temporal independence cannot be established from the supplied fields.

Run and download instructions cover the fixed archive and optional ucimlrepo export. Both download from UCI into ignored local storage. The client export records its own retrieval metadata and hashes; it is not silently substituted for the fixed archive. No RT-IoT records belong in Git.