Primary personalized-search checkpoints
Private recovery backup. Validation-selected trained checkpoints, original configs and all logged validation/test metrics. Code lives in Tyrec/rec_per, not this repo.
| Dataset | Checkpoint step | Validation-selected test |
|---|---|---|
| JDsearch | 5,000 | MRR@200 26.52% |
| KuaiSAR | 5,000 | NDCG@10 56.41% |
| AOL4PS | 10,000 | Full-catalog MRR@100 65.73% |
Recipes: causal Q/P, combined initial prefixes, stride-one overlap, all-query CE, intermediate Q weight 0.5, independent sigmoid fusion. HF dataset IDs and revisions are in each config. Test peaks differ from validation-selected scores; see metrics. JD/Kuai evaluation uses their released candidates; AOL uses all 951,940 documents. AOL frozen E5-small-v2 is referenced by revision in the config, not duplicated here. best.pt includes optimizer state as saved by training. Load only trusted checkpoints. Generated benchmark packs are batch-one, label-free requests for Tesla T4 timing; full indexes are cached vectors derived from these same weights, not model parameters. Amazon-M2 UK is backed up separately as Nithish2410/amazon-m2-uk-item-decoder-32m.
KuaiSAR checkpoint replacement
Run: rec_per_kuaisar_combined_per_query05_catalog_prefix_001.
Validation-selected step 5,000, also observed test peak. Same inputs, model,
all-Q loss and official candidates; only negative pool/exclusions changed:
100 fresh negatives/Q from the full catalog, excluding past/current-Q positives
including siblings. Future positives eligible; no validation/test-label exclusions.
latency/kuaisar/ still contains the previous step-3,500 benchmark pack. It is not
an inference index for this replacement checkpoint. Rebuild vectors from the new
weights before inference; never mix this checkpoint with that old cached index.
KuaiSAR primary update, 2026-11-02
Primary input contract: aol_timestamp_groups. Combined prefixes, stride-one Q windows,
partial oldest Q retained without its loss, timestamp-group positions and causal masks,
gap/session features, all-Q CE (historical weight0.5), split final positives,
independent-sigmoid query/history fusion. Selected by validation ndcg_at_5, not test peak.
Completed budget: 5,000 steps. Selected step: 5000.
Old primary preserved at archives/20261102_before_aol_style/kuaisar/27aa9f4a5cd8 and revision 27aa9f4a5cd8cedaa0174a9b8279f1d7991b5f26 (tag before-aol-style-kuaisar-20261102-27aa9f4a5cd8).
source_config.json pins the source dataset; config.json is the actual converted-data run config.
training-source.tar.gz preserves the exact experimental Python source. Active code now lives
in the personalization repository. Regenerate serving exports/catalog vectors from these weights.
Old latency/ packs remain historical and DO NOT match this primary. AOL4PS is unchanged.
Earlier replacement notes above are historical; this dated update and checkpoint_manifest.json are current.
JDsearch primary update, 2026-11-02
Primary input contract: aol_timestamp_groups. Combined prefixes, stride-one Q windows,
partial oldest Q retained without its loss, timestamp-group positions and causal masks,
gap/session features, all-Q CE (historical weight0.5), split final positives,
independent-sigmoid query/history fusion. Selected by validation mrr_at_200, not test peak.
Completed budget: 5,000 steps. Selected step: 5000.
Old primary preserved at archives/20261102_before_aol_style/jdsearch/72b6dc051370 and revision 72b6dc051370bc748998e7813c8b94c74aa950f0 (tag before-aol-style-jdsearch-20261102-72b6dc051370).
source_config.json pins the source dataset; config.json is the actual converted-data run config.
training-source.tar.gz preserves the exact experimental Python source. Active code now lives
in the personalization repository. Regenerate serving exports/catalog vectors from these weights.
Old latency/ packs remain historical and DO NOT match this primary. AOL4PS is unchanged.
Earlier replacement notes above are historical; this dated update and checkpoint_manifest.json are current.