tmlr-embedding-artifacts/recursive-geodesic-checkpoints

One Vector from Many: Hierarchical Long-Document Representation through Recursive Geodesic Transformer Composition

Research checkpoints for recursive composition of ordered document-chunk embeddings. The model combines a token-mass-weighted spherical reference with a learned bounded tangent correction and recursively reduces contiguous groups. The strict decoder predicts from the parent representation alone. These custom PyTorch checkpoints are not a Transformers AutoModel package.

Artifacts and roles

MANIFEST.json records exact filenames, original artifact roles, seeds, hosted paths, evaluation installation paths, byte sizes, SHA-256 hashes and metadata-only transformation mappings. SHA256SUMS independently lists the hosted checkpoint hashes.

Directory	Files	Role
final/	3	Final clean-v6 checkpoints used for final-model inference and the corresponding evaluations.
source/	3	Seed-matched continuation initialization checkpoints; not interchangeable with the final models.
historical_probes/	6	Historical fitted candidate/source parent-only probes, also included unchanged in the small supplement.
gradient_path/	6	Separately trained attached/detached decoder-parent-gradient intervention outputs; not replacements for the main final models.
decoder_probes/	45	Additional TRAIN-fitted direct/bounded/unbounded probe-family controls; distinct from the historical probes.
intermediate/	3	Shared warm intervention states. See each manifest entry's transformation record before using these as provenance artifacts.

Only verified, publishable artifacts are listed in the release manifest. An intermediate's metadata-sanitized derivative is a separate serialized artifact with its own hash; original experiment records are not rebound to that derivative. See the review source's checkpoint access guide for the exact transformation and any limitations.

Download and verify

Anonymous review source and reproduction instructions.

From that review checkout or the extracted supplementary ZIP:

python scripts/download_checkpoints.py --group final --verify
python scripts/download_checkpoints.py --group source --group historical_probes --verify

The standard-library helper uses the immutable 40-hex Hugging Face revision in reproducibility/HOSTED_CHECKPOINTS.json, sends no authentication credentials, downloads only the requested groups, verifies size/SHA-256, and installs exact evaluation protocol/lookup paths. Existing matching files and included probes are reused. --group all explicitly selects the entire collection. No token, gated-access approval or Hugging Face account is required.

For a manual Hub download from the same checkout, obtain the immutable release revision from the companion manifest, not from moving main:

REVISION=$(python -c 'import json; print(json.load(open("reproducibility/HOSTED_CHECKPOINTS.json"))["revision"])')
hf download tmlr-embedding-artifacts/recursive-geodesic-checkpoints \
  --revision "$REVISION" --local-dir checkpoint-files
cd checkpoint-files
sha256sum --check SHA256SUMS

The manifest inside this model repository describes the payload at the revision containing it; the companion source/ZIP manifest records that enclosing immutable revision. This avoids a self-referential commit hash in a file that would itself change that hash. Checksums detect byte changes; they are not authenticity signatures.

Separately versioned recovered caches

Six numeric cache bundles are an additive release, not additions to or replacements for the immutable 66-checkpoint inventory. CACHES_MANIFEST.json records all 8,075 extracted files, archive/member hashes and the two explicitly mapped metadata derivatives; CACHE_SHA256SUMS covers the six bundles and that manifest. The review source's reproducibility/HOSTED_CACHES.json pins their enclosing immutable revision separately from the weights.

Cache directory	Role
cache_kary4_strict_autoencoder_v6_clean_train	Original clean-v6 TRAIN leaves/teacher targets; original manifest matches the preserved historical checksum.
cache_kary4_strict_autoencoder_v6_clean_development	Clean-v6 development safe/full leaves and full-document indexes.
cache_kary4_strict_autoencoder_official_test_v6	Official safe/full leaves; explicitly identified anonymous manifest/COMPLETE derivatives.
cache_kary4_1024_expanded_source_seed424_v1	Earlier expanded-source TRAIN/validation leaves and targets.
cache_kary4_leaf510_vs1024_v1_leaf1024	Earlier leaf1024 TRAIN/validation tensors and sidecars; no original root manifest/COMPLETE was recovered.
cache_kary4_1024_validation3_strict_autoencoder_seed424_v2	Separate validation3 leaves and targets.

The bundles total 911,050,856 bytes compressed, 1,089,336,987 bytes extracted. All 8,032 PT files are unchanged originals. The official manifest alone replaces an identifying implementation-parent Git label with anonymous-source-snapshot-01; its COMPLETE points to the derivative manifest. Original/new metadata hashes are explicit. All tensor, index and sidecar bytes are preserved. Archive timestamps, ownership and permissions are neutralized.

All 8,032 tensor files loaded with weights_only=True on CPU and contained finite floating tensors. One real row per cache directory passed the published final model's recursive reducer and strict decoder. Complete static pickle/ZIP metadata and CRC inspection found no other tested identity/credential markers or recognized raw-text/token-sequence fields. These are scoped integrity/loadability checks, not a historical metric rerun, perfect privacy proof or full reproduction.

Do not bypass frozen guards. The official derivative is internally consistent but cannot satisfy the original manifest hash. Original config paths/hashes and some provenance hashes also differ from the normalized review layout. Frozen results/configs/guards remain unchanged. Some other study-specific vectors, tokens and task caches are still absent.

Upstream attribution and terms

These are numeric representations derived from the following datasets and pretrained teacher, not raw corpus text, token sequences or redistributed teacher weights. Dataset cards identify the licenses below; those labels do not settle every underlying web-page right or whether a particular derived representation triggers attribution/share-alike obligations. Retain applicable source terms; this project asserts no replacement blanket license.

Upstream creators/resource	Pinned source	Stated terms
Hugging Face FineWeb maintainers and source contributors	FineWeb, sample-10BT	ODC-BY 1.0
Hugging Face FineWeb-Edu maintainers and source contributors	FineWeb-Edu, sample-10BT	ODC-BY 1.0
OpenWebText maintainers and source contributors	OpenWebText	CC0 1.0
Wikimedia/Wikipedia contributors	Wikipedia, 20231101.en	CC BY-SA 3.0, GFDL, as stated by the pinned card
Alibaba Tongyi Lab	gte-modernbert-base	Apache 2.0

Document/source-index identifiers and content hashes remain intentionally available for provenance and can link cached representations to upstream documents. Embeddings can retain information about their inputs; no source-document privacy guarantee is made.

Scientific and reproduction limits

• The released final clean-v6 models use 300 decoder-only updates plus 100 joint updates from the exact continuation-source checkpoint, with a reinitialized decoder. They are not a from-scratch reconstruction. Recovered historical recipes/statuses trace the earlier source models through two encoder stages, but those earlier tensors and complete historical trainer/execution bindings remain missing; see the review source's partial lineage record.
• Restoring weights and the six recovered caches does not supply every pinned teacher/corpus/task input, study-specific token/vector cache, compatible software or GPU resource. Historical source-fingerprint discrepancies remain documented; no numerical result or old hash is rewritten to make a rerun pass.
• Separate older K-ary scaling checkpoints/probes and matched-pooling probes are outside the released 66-file inventory and remain unavailable unless explicitly listed in the companion access guide.
• The warm intermediates are not required to load the six final gradient-intervention outputs. Metadata-only copies do not satisfy guards requiring original serialized-byte hashes.
• Saved-score/statistical replay is not fresh model evaluation. Full-document roots have no direct full-root teacher target; retrieval results do not establish consistent superiority over pooling.
• Publication asserts no new license grant; consult the anonymous review source's terms. Public scientific titles, results and exact artifact hashes remain linkable to already-public research. This repository does not promise perfect anonymity against deliberate searches or platform operators.