zimplex/hrcf-big-results

HRCF big-model results

Completed HRCF training and storage/readback experiments requested by Kaisar Amangeldiyev. Amazon-Book is the priority result; Yelp is a backup. Both finished successfully on October 23, 2026, with all five evaluations and 15 storage variants per dataset. Amazon-CD was skipped.

Dataset	Embedding dimension	Parameters	Final Recall@20	Final NDCG@20
Amazon-Book	2048	191,836,160	0.15452681349268999	0.12260437130188574
Yelp	1024	118,986,752	0.09924957530436197	0.06715621294625061

Final clamped (cap1) and unclamped (cap0) metrics are identical for both datasets. The native JSONs retain full Recall/NDCG@5/10/20/50, evaluation histories, radius statistics, predictions and all storage variants.

Download

• hrcf_book_yelp_results_logs_20261024.zip: complete original results/ and logs/, plus their manifest.
• results/: eight unchanged native training, study, prediction and checksum files.
• logs/: five unchanged full logs, including both recipes, setup, GPU probe and launcher.
• summaries/: derived final-metric, five-point evaluation-curve and 30-variant CSV tables.
• FINAL_STATUS.json: completion, protocol and final metrics.
• MANIFEST.json: native file sizes, SHA256 and experiment coverage.
• CHECKSUMS.json: sizes and SHA256 for this published package. It excludes itself to avoid a circular checksum.

The ZIP and native JSON/log files retain their original bytes. CSV files are derived summaries. Per the collaborator's supplied README, large .pt/.ckpt files remain privately retained and are not included in this results handoff.

Protocol

HRCF upstream commit 452e9312ad041784be23224a9b4dafc7d43b6e3f, published settings, FP32 training, seed/split seed 1234, alpha 25, curvature 1, initialization scale 0.1, learning rate 0.002, weight decay 0.0005, batch size 10000, momentum 0.95 and one negative per example. Yelp uses four layers/margin 0.15; Book uses eight layers/margin 0.2. Each ran on one B200.

The configured budget is 500 epochs. The original published convention optimizes through epoch 499 and labels its final evaluation 500. Recorded evaluation labels are 100, 200, 300, 400 and 500. Both recipes exited zero and neither resumed or became non-finite.

After FP32 training, the supplied script emulates storage and readback in BF16, E4M3, E2M1, int8 and int4, under Lorentz, spatial and tangent representations. These 15 variants per dataset are scored from rounded/dequantized tensors in FP64. The script does not write packed low-bit tables or train low-bit models. The JSONs preserve published clamped scoring and unclamped ranking.

Book int4 predicted broken but was observed between; its native prediction_held=false is preserved. E2M1's broken prediction held. Full prediction thresholds, radius quartiles and representation-specific outcomes remain in the native files.

Source and resume correction

The completed runs used the collaborator's compatibility edits and a CPU checkpoint-load destination so RNG tensors can be restored on GPU. Training arithmetic/settings were unchanged, and both runs started fresh and completed uninterrupted. The later exact collaborator replacement adds sampler exception cleanup and first-resumed-epoch logging; it is installed for future resumes. Native sampler queues/streams are not checkpointed, so a future resume is not a bitwise continuation claim.

Full source, exact diffs, startup failures, native logs and review notes are in GitHub PR #1 and the pinned HRCF report.

Earlier HELM handoffs

The previously completed HELM results/checkpoints are available in their existing public repositories:

• HELM C4 20B checkpoints/results.
• HELM C4 2B low-precision checkpoints/results.