Log — Machine Learning Wiki

Append-only history. Entries start with ## [YYYY-MM-DD] <op> | <title> (split · ingest · query · lint).

[2026-07-30] split | machine-learning-wiki created from the hub _inbox (5 sources)

Created on the human’s instruction (“cluster them”) after the router, parking ml-system-design-case-studies, reported that ML strays had accumulated under four separate _inbox tags and that none of the hub’s 29 spokes owned machine learning itself.

Migrated from _inbox:

Why these five cohere. Every one of the five park records says the same thing in its own words — no spoke owns this — and each was declined by a different neighbour for a consistent reason: ../llm-inference-wiki is serve-time not train-time (both fine-tuning records), ../optimization-algorithms-wiki owns metaheuristics not gradient fitting (MLAlgorithms), ../platform-ops-wiki operates systems rather than designing them (the case studies). Four tags were four names for one absence. The five are three layers of one discipline: the fitting algorithm, the adaptation of a pretrained model, and the system built around it.

Two records deliberately left parked. ai-test-automation-paradox (ai-productionization, the other member of that tag) is AI-generated brittle tests — software-testing methodology, not shipping an ML system; folding it in would have widened the founding domain to “AI in engineering generally.” ai-engineering-from-scratch (learning-roadmaps) stays with its genre cluster on the substance-vs-genre argument the mlalgorithms park record already made: a curriculum is a meta-document, MLAlgorithms is the material. Adjacency notes updated on both.

Founding pages: 5 sources + 5 concepts (machine-learning, classical-ml-algorithms, fine-tuning, ml-system-design, demo-to-production-gap) + 1 tool (unsloth) + spine.

Entities. Created amd (Corporation) and kdnuggets (NewsMediaOrganization); linked the-new-stack cross-wiki from ../ai-governance-wiki rather than duplicating it (entity-index returned a 1.00 match). Authors — Kanwal Mehreen, Meredith Shubel, and the two GitHub maintainers — were kept as prose mentions rather than nodes: single-artifact contributors, below the relevance gate. Budget: 2 of 8 used.

Recorded as founding weaknesses, not discovered later. ai-projects-fail-infrastructure-people was re-fetched during the split and its body is still unreadable — WebFetch twice (including after a redirect to the canonical URL), the Firecrawl fallback, and a direct HTML fetch and parse all returned navigation and signup forms. It is ingested as an honest T4 stub carrying its headline and nothing more, per the hub’s never-drop-a-source rule. The corpus also has no middle layer (no end-to-end training source) and no independent benchmark — synthesis makes both the founding gaps rather than hiding them.

avoid-ai-writing run over the founding prose. Site rebuilt + verified at the clean rung, as the hub requires for a split.

[2026-08-02] ingest | Information Theory, Inference, and Learning Algorithms (MacKay, CUP 2003)

Routed here by the hub from Telegram — a book.pdf link with a Facebook click tracker on it, canonically inference.org.uk/itprnn/book.pdf. Downloaded (11.1 MB, 640 pages) and held at raw/mackay-itila.pdf; text extracted with pypdf for reading, since WebFetch can’t take a PDF that size. Runner-up spoke: none — see the note below on what nothing owns.

Tier T1, the spoke’s first. A Cambridge University Press textbook with derivations, rated exercises and worked solutions, by a researcher whose own LDPC results the book cites. Six sources now, and this is the only one that isn’t a vendor post, a catalog, an explainer or an unread headline.

Pages. Source summary information-theory-inference-learning-algorithms (Book), plus three concepts the corpus had no page for: bayesian-inference, information-theory, monte-carlo-methods. Updated machine-learning (a fourth thing under the three layers) and classical-ml-algorithms (its “no conceptual account is paged yet” line is now half-answered — the probabilistic branch only; still nothing on trees, boosting or SVMs, which the book simply doesn’t cover).

Entities. Created david-mackay (Person), cambridge-university-press (Organization) and university-of-cambridge (CollegeOrUniversity); entity-index returned no real match for any of them (the 0.75 hit on david-sacks is the fuzzy matcher catching a first name, not a person). Budget: 3 of 8. Shannon and Gallager are subjects of the book, not agents around it, so they stay prose mentions per ENTITIES.md.

Two things recorded rather than smoothed over. Part V is a 1990s Bayesian-neural-network account — single-neuron capacity, Hopfield nets, Boltzmann machines — written a decade before the scaling era, and the source page says so instead of letting “textbook” imply “current.” And the book opens the corpus’s first live seam: it says learning produces an ensemble of parameter values, while every other source here ships a point estimate and never mentions the difference. Filed in synthesis as open question 6, not as a contradiction — the two sides don’t cite each other.

Noted for the router, not acted on: Parts I, II, III and VI — data compression, noisy-channel coding, LDPC and fountain codes — are roughly half the book and no spoke in the hub owns information or coding theory. information-theory holds only what bears on learning, and says in a scope note that a real coding-theory body wants its own spoke rather than growth here.

avoid-ai-writing run over the new prose.

[2026-08-03] ingest | End-to-end forecasting with TimesFM 2.5 (MarkTechPost)

Routed from the hub (Telegram). New: timesfm-2-5-forecasting-tutorial (T3 source summary), timesfm (SoftwareApplication/model), time-series-forecasting (DefinedTerm/task), marktechpost (publisher). Updated: fine-tuning (the zero-shot limit case), demo-to-production-gap (a concrete sample of the contract details that decide it), synthesis (founding gap one-third closed + evidence-table row + new section + open question 7), index. Substance: google/timesfm-2.5-200m-pytorch, 200M-param transformer, zero-shot, ten quantiles per step (index 0 = MEAN, 1..9 = q10..q90, 5 = median), max_context to 1024 (multiple of 32), max_horizon 256 direct / recursive beyond, covariates via XReg in two fusion orders. The tutorial: synthetic retail panel (1,200 days x 6 stores, train 1,144 / holdout 56); MAE, RMSE, MAPE, sMAPE, MASE + pinball loss + 80% PI coverage against seasonal-naive and last-value; rolling-origin backtest (6 folds x 56 days, 36 store-fold combos); context ablation {64…1024} where accuracy plateaus and cost doesn’t (512 ~2x faster than 1024 at per_core_batch_size 16); anomaly detection by grading observations against the 80% interval (OK/WARNING/CRITICAL, ~20% flag rate on injected series); direct-vs-recursive long horizon; throughput sweep; NaN/short-context/ determinism robustness; production checklist. Why it matters here: the spoke’s founding gap explicitly named “no evaluation methodology” and this closes that third of it. The apparatus (MASE against a seasonal-naive baseline, rolling-origin folds, pinball loss + coverage, ablation against cost) is model-independent, which is why it counts as method. Second finding: timesfm is the pretrain-once pattern arriving in a non-text modality, and the limit case of the corpus’s “don’t fine-tune first” rule — not adapt later, but don’t adapt, in a field whose classical answer was a fresh ARIMA/ETS per series. T3 with the weakness recorded hard: all data is synthetic and the author generated it. Beating seasonal-naive on a series built from a weekly cycle plus noise shows the code runs, not that the model forecasts. Added as a third bias form in the evidence table (survivorship / first-party benchmark / synthetic problem). No ARIMA, ETS or GBM baseline; no independent replication. Entities: marktechpost paged as publisher; Google reused cross-wiki (google, llm-providers-wiki) rather than duplicated, per ENTITIES. backtesting cross-linked to quant-trading-wiki as the trading-side sibling. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-08-04] ingest | Transformers-Tutorials (NielsRogge)

Routed here by the hub (runner-up: none — no other spoke covers model training). T2, the same grade and shape as mlalgorithms: a personal MIT teaching repository, not a vendor artifact. The affiliation is recorded rather than penalized — Rogge contributed many of these model implementations to the Transformers library, which makes him expert on the code and an interested party on the choice of library. Nothing here compares Transformers to an alternative.

New pages (2): transformers-tutorials (source), niels-rogge (Person). Updated: fine-tuning context via synthesis, index, growth edge 1.

Closes the second third of the founding gap. The corpus jumped from how a fitting algorithm works to how to adapt a finished model, with nothing on training end to end. timesfm-2-5-forecasting-tutorial supplied the evaluation methodology on 2026-08-03; this supplies the training loop — executable rather than described, fine-tuning on real datasets across several dozen models, written three ways (native PyTorch, Lightning, Trainer) with a stated basis for choosing. fine-tuning had been concepts from an explainer with no measurements; unsloth hides the loop. This shows it.

Remaining hole is now specific: data collection and labelling, plus pretraining proper.

The pattern of closure is itself the finding. Both thirds arrived as method carried incidentally by teaching material — one inside a tutorial for a single model, one inside a tutorial collection for many — never as a source about the training pipeline as a subject. The middle layer appears to have no literature of its own within this corpus’s reach; it gets documented only where someone was demonstrating something else. Folded into “The founding gap” rather than opened as a dated section.

No numbers, and none claimed — a teaching resource, so the spoke’s record-the-setup rule resolves trivially. Growth edges 2 (independent benchmark) and 3 (evaluation as a discipline) untouched.

[2026-08-05] ingest | Supervision (Roboflow)

Routed from the hub. supervision — Roboflow’s model-agnostic Python computer-vision utility library, MIT, 48,725★, T2. Created roboflow (Corporation; entity-index returned no match ≥ threshold). Budget: 1 of 8.

Module structure read from the tree, not the README — 78 modules under supervision/: detection (the sv.Detections structure, NMS, zones), annotators/draw, tracker, dataset (YOLO/COCO/ Pascal VOC conversion), metrics (mAP), plus classification, keypoint, geometry, validators.

This ingest stretches the spoke’s scope boundary and the page says so. CLAUDE.md declares train-time and design-time; nothing in this library runs during backprop. It was routed here because the alternatives fit worse — llm-inference-wiki is serve-time mechanism for LLMs (logits, sampling, KV cache), and dev-tooling-wiki declines application-level runtime libraries under the Ky precedent. Logged as a stretch, not a precedent: if more serve-time vision tooling arrives, either widen the boundary on purpose or admit that computer vision as a modality is unowned hub-wide.

Gap-relevance is real rather than nominal. Growth edge 1’s live half is data curation and evaluation, and this source ships both as installable modules — so the edge note now distinguishes holding them as tooling from holding them as a result, which the corpus still does not.

New synthesis section on demo-to-production-gap: the corpus held that gap as a claim with no mechanism (a success catalog with no denominator, an unreadable failure headline). This itemizes it — format conversion, identity across frames, drawing a box, counting line crossings, computing mAP. Weak evidence in one direction only (a library’s existence proves demand, not that projects fail without it), and stated that way on the page.

Vendor-adjacency recorded: Roboflow sells a CV platform and the library feeds its inference product. No performance or accuracy claims anywhere in the repo, so nothing to verify.

[2026-08-05] ingest | SkillSmith

Routed from the hub (runner-up: agentic-tooling-wiki, declined on the homonym — see below). skillsmith — arXiv preprint “Learning to Compose Parametric Skills and Textual Knowledge” (Dery, Tjandra, Samiei, Kuncoro, Yahav, Shen, Szlam), T2.

The spoke’s first primary research on adaptation (prior adaptation sources were an explainer and a toolkit). Method: prefix-tuning stores a skill as weights; an LLM reads existing prefix weights + text and emits new prefix weights for a target skill; reported to beat text-only and weight-space-only baselines. Prefix-tuning is PEFT, so it sits in the fine-tuning family — updated that page with the “adapter as an object you compose, not just train and load” reframing.

Synthesis: added as a third stance on the founding weights-as-distribution vs point-estimate seam — it treats trained weights as manipulable objects, making parameters the unit of analysis. Held loosely: preprint, authors’ own evaluation, benchmark caveat in full force.

Homonym recorded: a parametric skill (prefix weights) here vs an agent skill (markdown prompt) in agentic-tooling-wiki — both senses arrived in the hub the same day (education-agent-skills there). That homonym is exactly why agentic-tooling was declined: “skills” in the title is weight-space, not prompt-space, so this is model-adaptation research and files here by subject.

Entities: the 7 authors are deferred — a preprint with no affiliations recoverable from the abstract; thin evidence, and paging seven one-off Person nodes would breach the evidence-only rule. Page the lead author on recurrence or if a query needs it.

[2026-08-07] ingest | Phase Transitions in Attention: A Bayesian Theory of Copy Head Emergence

Routed from the hub (Telegram). arXiv 2606.12058v1, Lavie & Fischer et al., 10 June 2026 — T1, and a preprint, so no peer review. The spoke’s first primary research on training itself rather than on a tool, a tutorial or a finished model.

New pages: phase-transitions-in-attention (source summary), capability-emergence, induction-head. Updated bayesian-inference with the section that source earns.

Why it matters here. Open question 6 — does the Bayesian account survive at scale — had a textbook on one side and a point-estimate toolchain on the other, neither citing the other. This paper puts them in one equation for a single attention head: the analytically-derived posterior’s MAP is where Adam lands (minimising the action at large χ = gradient flow with weight decay), and SGLD agrees. What makes it computable is a permutation symmetry, not a scalable method, so the question is now sharper rather than closed. Recorded as a partial answer, and a new open question 8 for the monitoring claim.

The finding that travels furthest is not about training: transition order depends on the attention activation. Softmax jumps first-order with no precursor signal; linear attention goes second-order then crosses over gradually. The paper’s own reading is that capability onset in softmax models is “intrinsically unpredictable from observable quantities,” with “fundamental limits to what can be monitored.” Held as a conjecture with a derivation behind it — one layer, one head, i.i.d. uniform tokens, supervised copy task, equivalent-kernel and d_model → ∞ approximations, and copying is the minimal circuit under in-context learning rather than in-context learning itself.

The PDF would not extract through WebFetch (raw object stream); pulled the text locally with pypdf.

Entities: 3 created (itay-lavie, zohar-ringel, moritz-helias), 0 linked. Six authors and five affiliations; I paged the corresponding author and the two senior authors whose groups the work joins, and left Fischer, Lekov and Van Maele as plain mentions along with all five institutions. Same call as the hub’s 2026-08-06 precedent on thin one-off paper authors — revisit if this line of work recurs, since Fischer is a co-first author and would be the next node.

[2026-08-08] ingest | Data Cascades + Pervasive Label Errors — the founding gap’s last third

Two peer-reviewed sources (both T1), arrived via the research pass hunting growth edge 1, the data-collection-and-labelling half of the missing middle layer.

data-cascades — Sambasivan, Kapania, Highfill, Akrong, Paritosh, Aroyo, CHI 2021, google Research. 53 practitioner interviews across India, East and West African countries and the USA, all in high-stakes domains. 92% had hit at least one compounding data failure; 45.3% two or more in one project. Cascades are “pervasive, invisible, delayed, but often avoidable.”

pervasive-label-errors — Northcutt, Athalye, Mueller, NeurIPS 2021. Ten standard test sets, confident learning plus crowdsourced confirmation: ≥3.3% of labels wrong on average, ≥6% of ImageNet validation, 51% of algorithmically flagged candidates confirmed. Correcting the labels reverses model rankings — ResNet-18 over ResNet-50, VGG-11 over VGG-19.

New pages: data-cascades, pervasive-label-errors, training-data-quality (the concept node consolidating both), nithya-sambasivan, curtis-northcutt. Entity budget 2 of 8; google linked cross-wiki rather than duplicated.

A correction to synthesis, not just an addition. This wiki had written that the middle layer “appears to have no literature of its own in this corpus’s reach” and is documented “only where someone was demonstrating something else.” That was wrong. The literature exists in HCI and in the datasets-and-benchmarks track; it studies practitioners and labels instead of architectures, so ML-side searching never returned it. The old text is kept and the correction written under it.

Contradiction recorded, not resolved. ai-projects-fail-infrastructure-people (T4, body never readable) says systems fail on infrastructure and people; data-cascades (T1, 53 interviews) says the cause sits in data decisions taken before any model exists. Both kept on demo-to-production-gap. Also flagged there: 92% is prevalence among practitioners recruited to discuss data practice in high-stakes domains, not a failure rate for AI projects, so the corpus’s open question about a denominator stays open.

Edge 2 moves rather than closes. The label-error result means a “neutral third-party benchmark” — the strong rung on this spoke’s evidence spectrum — inherits its test set’s error rate. The edge now asks for a benchmark that states its label quality.

Successor edge 1a written: nothing here says what competent labelling looks like — no annotation guidelines, no inter-annotator agreement, no sampling design, no cost.

[2026-08-08] ingest | Measuring Annotator Agreement — the instrument for the data layer

Research pass (20-edge run), successor edge 1a, written hours earlier the same day. measuring-annotator-agreement new (T1, Braylan/Alonso/Lease, arXiv 2212.09503, ACM 2022).

Completes a sequence rather than adding to a pile: data-cascades says the practice fails (92% of practitioners), pervasive-label-errors says the output is wrong (≥3.3% of labels), and this says the instrument you would check with is well-founded only for the simplest label shapes. Agreement across bounding boxes, keypoints, sequence tags, ranked lists, free-text translation, numeric vectors and syntax trees needs a distance function chosen first, and the metric inherits that choice.

Cross-corpus rhyme noted on the page: er-evaluation-framework in research-wiki found the naive entity-resolution metric biased upward. Different field, same shape — the error hides in the measurement of the measurement.

Edge 1a marked partly closed: the corpus now has a metric and still has no documented labelling practice — no guidelines, no annotator training, no sampling design, no cost.

[2026-08-09] ingest | Gradient-boosted trees — XGBoost and the tabular benchmark (via research pass)

Coverage edge 4 closed with exactly the pair it named. xgboost — Chen & Guestrin, KDD 2016, T1, extracted from the PDF locally. tabular-deep-learning-benchmark — Grinsztajn, Oyallon & Varoquaux, NeurIPS Datasets & Benchmarks 2022, T1. Concept page: gradient-boosted-trees.

What was missing. classical-ml-algorithms carried the entire non-deep story, and the method practitioners reach for first on tabular problems had no page — while open question 7 was explicitly asking for it as the baseline a time-series foundation model should be measured against.

Two distinct claims, recorded separately. XGBoost’s contribution is a system: the algorithm was Friedman’s, and the paper’s own numbers are engineering ones — 0.68 s per tree against scikit-learn’s 28.51 s at the same AUC on Higgs-1M, 1.7 billion Criteo examples processed on a single machine. The 17-of-29-Kaggle-winners figure is adoption evidence from a venue with its own selection effects, and is labelled as such.

Grinsztajn et al. is the claim that ages well: 45 datasets, ~20,000 compute hours of tuning per learner, tree ensembles still state-of-the-art at ~10K samples, with three named inductive biases behind it — robustness to uninformative features, orientation preservation against rotation-invariant networks, and irregular target functions. That is a structural explanation, so it predicts rather than merely reports.

Open question 7 is unchanged in status and better specified: the baseline now exists in the corpus, the comparison still does not. Edge 4’s successor asks for it directly.

Entities: 0 created — no new author or org pages; one paper each is not a recurrence.

[2026-08-10] ingest | Awesome Agentic Time Series — 243 papers, and the shape of them

Routed here by the hub (runner-ups: agentic-tooling-wiki for the agent-architecture vocabulary, quant-trading-wiki for the finance slice). Source: https://github.com/TROUBADOUR000/Awesome-Agentic-Time-Series, T3, MIT, 188★/10 forks, created 2026-06-01, read 2026-08-10. Companion to an unpublished survey, The Landscape of Agentic Time Series Systems.

Pages: awesome-agentic-time-series. time-series-forecasting gained a section on the field around it.

Why this spoke. The subject is time-series analysis, which this wiki already owns (time-series-forecasting, timesfm). The agents are the method. That follows the hub’s standing precedent — a tool that merely uses LLM agents files by subject, with the mechanism cross-linked — and the list’s own domains settle it: weather, IIoT, motion and multimodal classification outnumber finance, so quant-trading-wiki would have held a tenth of it.

The finding is the distribution, not any entry. I counted the sections rather than reading the papers: 133 papers on building agentic time-series systems, 16 on evaluating forecasts or agentic decisions, 7 on reliability. Under a tenth of the field is about whether any of it works. That is demo-to-production-gap arriving from the literature side, and it lands directly on time-series-forecasting‘s standing complaint that nothing here compares a pretrained forecaster against classical baselines — the comparison is the field’s thinnest corner, not this wiki’s blind spot.

Also recorded: 31 foundation-model entries (Falcon-X, Toto 2.0, Timer-S1 at billion scale, TS-Memory) mean the pretrain-once move in this modality is crowded, and timesfm is no longer the whole story here.

T3 and honestly so. A three-month-old maintainer-curated list with no published inclusion criteria, a “200+” badge that actually counts to 243, and a BibTeX entry with no venue or year. Nothing in it was read; every claim about a paper is that paper’s title. The counts are what this page stands behind.

[2026-08-10] ingest | Who Annotates in NLP? (research pass)

Arrived through the hub Research Pass, against ## Most wanted #3 — a documented labelling operation, the edge left half open on 2026-08-08 when measuring-annotator-agreement supplied an instrument and no practice.

who-annotates-in-nlp written (T1, arXiv 2606.02255; preprint, flagged as such). It audits 2,667 annotation tasks from 1,603 ACL-venue papers and validates its own LLM extraction against a 72-task adjudicated gold standard at Krippendorff’s α 0.606, above the 0.585 human–human figure on the same data.

The edge does not close, and the reason is the finding. The practice this wiki kept failing to locate is largely unpublished: annotator training reported for 18.7% of tasks, released guidelines 34.1%, an agreement value 35%, against recruitment at 90.4%. The search was not at fault; the base rate is. Re-specified on the hub shortlist rather than struck.

Placed against data-cascades, pervasive-label-errors and measuring-annotator-agreement — the four now say, in order: the damage is real, the labels are wrong at measurable rates, the instrument is ill-suited to most label shapes, and the conditions needed to interpret the instrument are usually missing.