GwendalTsang/bwdam-icml-2026-reproduction-bundle

BW-DAM ICML 2026 reproduction bundle

This dataset repository is the standalone reproduction bundle for Dense associative memory for Gaussian distributions, ICML 2026 paper #14930. The challenge statements match arXiv:2509.23162v1, not the subsequently changed v2 settings, so every result records the v1 version lock.

The bundle contains executable code, deterministic tests, raw per-seed CSV/JSON, aggregate statistics, digitized figure references with uncertainty, RTX 3070 results, Jetson Orin audit logs, a hash-verified T4 manifest, Text8 trainer/source-control metadata, and the interactive reproduction poster. It is an experiment archive rather than a row-oriented benchmark dataset.

Claim verdicts

Claim	Verdict	Reproduced evidence
1. Wasserstein LSE	Verified	Euclidean/common-covariance reductions agree to at most 2.17e-15
2. Exponential capacity	Supported under v1 assumptions	10/10 separation trials pass at every nontrivial d=64..192; this checks the theorem schedule, not an empirical failure threshold
3. Exponential retrieval error	Supported with finite-d caveat	RTX slope -1.70923/d, 95% bootstrap CI [-1.75951,-1.65095]; Theorem-2 contraction constant first becomes <1 analytically at d=325
4. N=10000,d=50 temperature separation	Verified for Figure 3	beta 1 retrieves perfectly; beta 0.1 final mean W2 7.233475 ± 0.000124; Figure 6 uses N=5000,d=25
5. Text8 phase transition	Partially reproduced / trainer-sensitive	Five batched-objective models give 10/50/90% crossings 9.495/13.121/15.638; literal/corrected pinned-source controls give 28.079/41.201/62.243 and 25.353/36.466/56.284
6. Non-commuting covariances	Verified at theorem radius	beta 1 one-step accuracy is 100% across five seeds at r; at 100r, accuracy is 0.9573 and Figure 8 itself plateaus near 0.16, contradicting its prose

Repository map

Path	Contents
bwdam.py	Bures--Wasserstein distances, transport/update rules, v1 samplers, exact perturbations
experiments.py	Standalone CLIs for Claims 1--6, raw trials, confidence intervals, plot/raster comparisons
test_bwdam.py	20 deterministic formula, geometry, chunking, and log-domain checks
hf_job.py	Standalone staged Hugging Face GPU Job entry point
train_word2gauss.py	Deterministic batched Text8 Word2Gauss-objective trainer
train_word2gauss_source_control.py	Hash-gated literal/corrected single-worker Cython source control
results/local/, results/rigorous/	RTX results and refined figure comparisons
results/hf-t4/	downloaded main-account T4 results and manifest
rigorous_data/	training metadata and, where included, source-control checkpoints
results/rigorous/jetson_audit/	hardware inventory, source hashes, test logs, tegrastats, Claim-3/4 outputs
poster/	Posterly HTML/embed, preview, figures, and validation reports
PROVENANCE.md	immutable inputs, exact hashes, hardware, Jobs, and caveats

The canonical Text8 checkpoints are mirrored separately in the model repository. The 31,344,016-byte Text8 archive is intentionally not redistributed.

Download and verify

hf download GwendalTsang/bwdam-icml-2026-reproduction-bundle \
  --repo-type dataset --local-dir bwdam-reproduction
cd bwdam-reproduction
python3 -m venv .venv
. .venv/bin/activate
python3 -m pip install -r requirements.txt
python3 test_bwdam.py

Expected result: 20/20 tests pass. CUDA-enabled PyTorch is needed for the reported paper-scale timings; the deterministic formula suite also runs on CPU.

Rerun Claims 1--4 and 6

python3 experiments.py --device cuda --dtype float64 \
  --backend rerun --output-dir rerun_outputs claim1
python3 experiments.py --device cuda --dtype float64 \
  --backend rerun --output-dir rerun_outputs \
  claim2 --dimensions 32 64 96 128 160 192 --trials 10
python3 experiments.py --device cuda --dtype float64 \
  --backend rerun --output-dir rerun_outputs \
  claim3 --dimensions 64 96 128 160 192 --trials 10 \
  --queries 256 --bootstrap-draws 2000 --control-beta 0.001
python3 experiments.py --device cuda --dtype float32 \
  --backend rerun --output-dir rerun_outputs \
  claim4 --trials 5 --q-chunk 256
python3 experiments.py --device cuda --dtype float64 \
  --backend rerun --output-dir rerun_outputs \
  claim6 --trials 5 --q-chunk 25 --n-chunk 250

Claim 6's full four-configuration run takes several minutes on an RTX 3070 because it uses general covariance eigensolvers in float64. Do not silently substitute the commuting diagonal path.

Rerun Claim 5 from published checkpoints

hf download GwendalTsang/bwdam-text8-word2gauss-50d \
  --local-dir published_checkpoints
python3 experiments.py --device cuda --dtype float32 \
  --backend rerun --output-dir rerun_outputs claim5 --query-trials 2 \
  --embeddings \
    published_checkpoints/text8_word2gauss_v1.npz \
    published_checkpoints/text8_word2gauss_seed17.npz \
    published_checkpoints/text8_word2gauss_seed29.npz \
    published_checkpoints/text8_word2gauss_seed41.npz \
    published_checkpoints/text8_word2gauss_seed53.npz

Evaluate source controls separately:

python3 experiments.py --device cuda --dtype float32 \
  --backend literal-source-rerun --output-dir rerun_outputs claim5 \
  --query-trials 2 \
  --embeddings published_checkpoints/text8_word2gauss_source_literal.npz
python3 experiments.py --device cuda --dtype float32 \
  --backend corrected-source-rerun --output-dir rerun_outputs claim5 \
  --query-trials 2 \
  --embeddings published_checkpoints/text8_word2gauss_source_corrected.npz

For fresh batched training, train_word2gauss.py downloads Text8 when absent and checks the printed hash:

python3 train_word2gauss.py --device cuda --seed 17 \
  --backend rerun --output data/text8_word2gauss_seed17.npz

Required Text8 SHA-256: a6640522afe85d1963ad56c05b0ede0a0c000dddc9671758a6cc09b7a38e5232.

For literal source training, clone and check out the exact public commit, build it with Cython 3.2.4 and NumPy 2.5.1 using --no-build-isolation, then run:

python3 train_word2gauss_source_control.py \
  --source-dir /absolute/path/to/word2gauss \
  --variant literal --text8-zip data/text8.zip \
  --output rigorous_data/text8_word2gauss_source_literal.npz \
  --full-checkpoint rigorous_data/text8_word2gauss_source_literal.tar.gz \
  --metadata-output rigorous_data/text8_word2gauss_source_literal.json \
  --epochs 5 --chunk-tokens 100000 --backend literal-source-rerun

The runner aborts unless embeddings.pyx matches the expected hash. For the corrected control, change only sigma_ptr[j] to sigmaj_ptr[0], confirm source SHA-256 56e247f63a4dbd0168d5dcc204fe6b8fe01da93bce3629cb8c8513f9e7e183f2, rebuild, and use --variant corrected with corrected output names.

Reproduce the cloud check

First stage this flattened bundle plus data/text8_word2gauss_v1.npz below the source/ prefix of the run Bucket. Then launch:

hf jobs run --detach --namespace GwendalTsang \
  --name bwdam-rigorous-crosscheck-corrected \
  --flavor t4-small --timeout 30m \
  --volume hf://buckets/GwendalTsang/repro-dense-associative-memory-for-gaussian-distributions-runs/source:/workspace/repro_bwdam:ro \
  --volume hf://buckets/GwendalTsang/repro-dense-associative-memory-for-gaussian-distributions-runs:/outputs:rw \
  pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime \
  python /workspace/repro_bwdam/hf_job.py rigorous

The completed Job ran 389 billed seconds on t4-small, approximately $0.0434 at $0.0067/min. Its five subprocesses (20 tests and Claims 3--6) all exited zero, and all eight result hashes were verified after Bucket download. The internal program interval was 382.77 seconds with estimated compute $0.04274.

The superseded preflight is retained for auditability: it ran 17 seconds (approximately $0.0019), passed the tests and Claim 3, then exited 2 because Python 3.11 treated the exact subcommand flag --d as an abbreviation of top-level --device/--dtype. Setting allow_abbrev=False fixed the cross-version parser behavior. It contributes no claim evidence.

Jetson portability evidence

The independent device was a Jetson Orin Nano Developer Kit (Ubuntu 22.04.5, CUDA 12.6, PyTorch 2.11.0, about 7.98 GB unified memory). It passed all 20 CPU tests and dedicated commuting CUDA checks. Claim 3 gave log10-error slope -1.69582, 95% bootstrap CI [-1.79638,-1.64500]; Claim 4 reproduced perfect beta-1 retrieval and beta-0.1 failure over three seeds.

Its installed generic batched float64 torch.linalg.eigh path failed with CUSOLVER_STATUS_EXECUTION_FAILED, so Claim 6 was not run on Jetson CUDA. The bundle retains this environment failure verbatim and uses RTX float64 for the non-commuting evidence.

jetson_runner.py requests credentials interactively and never stores them. A typical probe/sync sequence is:

python3 jetson_runner.py --host JETSON_HOST --user JETSON_USER probe
python3 jetson_runner.py --host JETSON_HOST --user JETSON_USER sync

Version and integrity anchors

• arXiv v1 source archive SHA-256: a88a9b572b0d3c28b62de7f9419885cd3367d25e8b453bf4ffeba36a68f1f6ee
• v1 main_arxiv.tex SHA-256: 9bbcddf2cbbd433bb4c5928281c462c01fa45069d68c175e815f6d6190b72b2e
• Text8 ZIP SHA-256: a6640522afe85d1963ad56c05b0ede0a0c000dddc9671758a6cc09b7a38e5232
• Word2Gauss commit: cdf5e7f5d0c8c7582fdf92e4599bd8efda5b98bc
• literal/corrected embeddings.pyx SHA-256: a9c00384...7d8e / 56e247f6...83f2
• main T4 manifest SHA-256: abbc5e2f5e897513530964e4ebf1dffc3431ef444109ba526cbe8b112eb7a72f

See PROVENANCE.md for full hashes, figure rasters, hardware, checkpoints, and result manifests.

Licensing and intended use

No umbrella license is asserted for this mixed-provenance archive. Reproduction code written for this work, third-party Word2Gauss source, Text8-derived checkpoints, paper figures/digitizations, and generated poster assets do not necessarily share the same terms. Word2Gauss code is MIT-licensed, but Text8 is not redistributed and no license for Text8-derived model parameters is inferred. The YAML therefore deliberately omits license.

Use the bundle to inspect and rerun the scientific claims. Do not treat the word embeddings as validated production representations or use them for decisions about people.

Linked resources

• Canonical logbook Space
• Text8 checkpoint model
• Persistent Job-output Bucket
• Trackio artifact Bucket
• Private agent-trace dataset
• Resource collection
• Completed T4 Job
• Superseded failed preflight
• Exact public Word2Gauss commit