fromthesky/pldr-llm-training-dynamics-data

PLDR-LLM Training Dynamics Data

Reported numerical evidence for Training and Inference Dynamics of PLDR-LLMs: Row-Map Collapse, Renormalization, and Predictive Reduction, by Burc Gokden.

• Monograph: Huggingface Paper Page.
• Scientific code and readers: GitHub repository.
• Numerical evidence: Hugging Face dataset.
• Citation: CITATION.cff.

This compact dataset contains numerical records, complete reported outcome grids, controls, uncertainty summaries, statistical roles and portable provenance. It includes no manuscript source, PDF, executable helper, corpus text, token array, model checkpoint or large raw observation array.

Access

Use Python 3.11 or later and the readers in the GitHub companion. These commands run from the code repository; DATA-REPO is the location of this dataset clone.

python3 scripts/verify_evidence.py --data-repo DATA-REPO
python3 scripts/read_evidence.py --data-repo DATA-REPO --list
python3 scripts/verify_evidence.py --data-repo DATA-REPO --extract /tmp/pldr-evidence

The verifier authenticates all indexed compressed objects, their decompressed bytes and the display/claim links. No private research repository is needed.

Contents and interpretation

• index.json maps logical scientific identities to deterministic gzip objects in objects/, with compressed/uncompressed SHA-256 hashes and byte counts.
• coverage.json indexes all table and figure environments by monograph label, final number and chapter subject, along with mathematical statements and conditional empirical claim scopes.
• displays/ logical records transcribe printed table cells at their displayed precision and figure captions/axis annotations. Related numerical-record links aid navigation. They are subject associations, not a claim that the complete raw plotting dependency graph or every underlying point is deposited.
• checks/cache-reconstruction/ logical records give the bounded raw-array reconstruction result, analysis and independent verification. They do not include the external raw arrays or claim native model replay.
• data-dictionary.json records observables, units, centering, denominators, independent statistical units, nested measurements and pairing.
• raw-inputs.json identifies separate raw-array and native replay inputs by portable acquisition role. Large input hashes also occur in protocol records.
• normalization.json binds exported records to acquisition-record content hashes and specifies the metadata policy. Local execution locations and diagnostic environment strings are omitted or replaced by logical roles. Device identifiers and acquisition dates are omitted. Elapsed durations, training-time coordinates, seeds and source positions remain available.
• manifest.json, SHA256SUMS and validation.json bind the prepared dataset and record the numerical-equivalence and coverage checks.

Families row, rg and model identify row dynamics, chronological renormalization and model-wide dynamics. They do not imply independent datasets or replications. Numerical values, signs, uncertainties, seeds, counts, source positions and outcomes are preserved. Normalized metadata changes record bytes; new object hashes are provided. A normalized protocol documents an acquisition but does not replace the immutable protocol required for raw reanalysis.

The 30-cell cache confirmation retains eighteen recalibrated aggregate passes, six zero-control transfer passes and six positive-control initial-cache transfer failures, including context exceptions. Other failed precision gates, rejected reductions, null controls and finite-scope qualifications remain reported.

Single-pass consumption of distinct registered RefinedWeb target blocks is the primary training law. Repeated-corpus controls remain separate. Heads, layers, contexts and times are nested or paired; they are not extra independent model replicas. Source-selection records include identifiers, hashes and positions, not corpus contents. Consult each protocol and the monograph for conditioning, calibration/assessment roles and uncertainty conventions.

This is a reported-evidence deposit. Raw-array reduction and native replay need the separately identified assets and the code companion's admission checks. Integrity checking and display transcription do not constitute new scientific replication. Rights and attribution remain with their respective materials; this deposit makes no blanket third-party data-license grant.

Published PDF destinations

The mathematical entries in coverage.json retain the statement's stable id, printed number, original-build anchor, and shared counter. Statement aliases and nested equation/clause labels remain separate. For navigation in the published monograph, use published_anchor and published_pdf_page. The latter is a one-based PDF page index, including front matter.

The top-level published_pdf record identifies arXiv 2609.34130v1, its versioned PDF URL, SHA-256 and page count. All 263 numbered statements have a published destination. The original-build anchors are retained for provenance; 212 of them differ from the published destinations. This mapping concerns mathematical statements; table/figure coverage and numerical records retain their existing meanings.

These entries agree with the code companion's provenance/statement-manifest.json. Its check_formal_manifest.py --data-repo check verifies code/data correspondence using the standard library. To check the actual PDF destinations and pages, run from the code repository:

python3 -m pip install -r requirements-publication.txt
curl -fL https://arxiv.org/pdf/2609.34130v1 -o /tmp/pldr-monograph-2609.34130v1.pdf
python3 scripts/check_published_pdf.py --pdf /tmp/pldr-monograph-2609.34130v1.pdf --data-repo DATA-REPO --output validation/published-pdf.json

The PDF checker authenticates the supplied PDF before resolving its destinations. It uses the separately pinned pypdf package; the evidence reader continues to need only the standard library. Use corresponding code and dataset revisions that include the published-PDF mapping.

Citation

CITATION.cff provides the preferred monograph citation with the arXiv v1 identifier. Also record the dataset Git revision and the manifest.json payload SHA-256 used. These identify the evidence release separately from the manuscript version. Coverage numbers refer to the monograph.