| --- |
| license: other |
| tags: |
| - recommendation |
| - generative-retrieval |
| - semantic-ids |
| --- |
| |
| # OneDiffRec semantic-ID codebooks |
|
|
| Item semantic-ID (SID) tables and the trained tokenizers behind them, from the OneDiffRec |
| grid on Amazon Reviews (Industrial_and_Scientific and Office_Products). Three tokenizers, |
| each at 3/4/5 codebooks x 128/256/512 codes: |
| |
| - `MQ`: TokenRec parallel multi-codebook tokenizer (rq/mq.py, 10000 epochs, seed 2024), the |
| fixed implementation (commit 7ae3512, tables rebuilt 2026-10-16). |
| - `rqkmeans`: residual k-means. |
| - `rqvae`: RQ-VAE (rq/rqvae.py, 10000 epochs, seed 2024). |
| |
| ## Layout |
| |
| ``` |
| <dataset>/tokenizers/tables.json settings used to build all tables |
| <dataset>/tokenizers/<method>_<D>codebook_<K>/ |
| <Dataset>.*index*.json native SID table: item index -> ["<a_i>", "<b_j>", ...] |
| done.json build record: collisions, sha256, geometry source |
| geometry/ last-codebook centroids, residual vectors, item ids |
| (MQ: also mq_model.pth, the trained tokenizer) |
| checkpoints/.../best_collision_model.pth RQ-VAE only: the model the table comes from |
| zcr-*.index.json rqkmeans only: zero-collision table |
| tokenizer.log, zcr.log |
| <dataset>/zcr/<method>_<D>codebook_<K>/ |
| index.json zero-collision-repair (ZCR) SID table |
| native_codes.npy, zcr_codes.npy, native_index.json, item_ids.json |
| final_centroids.npy, last_stage_vectors.npy, manifest.json, stats.json |
| ``` |
| |
| Native tables let several items share one SID; ZCR moves colliding items to the nearest free |
| last-level code so every item has a unique SID. Absolute paths inside manifests and logs |
| refer to the original training cluster. |
|
|
| ## Configurations |
|
|
| | dataset | method | codebooks | codes | items | items sharing a SID | ZCR table | |
| |---|---|---|---|---|---|---| |
| | industrial | MQ | 3 | 128 | 3105 | 1736 (55.9%) | yes | |
| | industrial | MQ | 3 | 256 | 3105 | 1321 (42.5%) | yes | |
| | industrial | MQ | 3 | 512 | 3105 | 1048 (33.8%) | yes | |
| | industrial | MQ | 4 | 128 | 3105 | 1181 (38.0%) | yes | |
| | industrial | MQ | 4 | 256 | 3105 | 893 (28.8%) | yes | |
| | industrial | MQ | 4 | 512 | 3105 | 735 (23.7%) | yes | |
| | industrial | MQ | 5 | 128 | 3105 | 812 (26.2%) | yes | |
| | industrial | MQ | 5 | 256 | 3105 | 629 (20.3%) | yes | |
| | industrial | MQ | 5 | 512 | 3105 | 591 (19.0%) | yes | |
| | industrial | rqkmeans | 3 | 128 | 3105 | 1372 (44.2%) | yes | |
| | industrial | rqkmeans | 3 | 256 | 3105 | 975 (31.4%) | yes | |
| | industrial | rqkmeans | 3 | 512 | 3105 | 514 (16.6%) | yes | |
| | industrial | rqkmeans | 4 | 128 | 3105 | 873 (28.1%) | yes | |
| | industrial | rqkmeans | 4 | 256 | 3105 | 624 (20.1%) | yes | |
| | industrial | rqkmeans | 4 | 512 | 3105 | 302 (9.7%) | yes | |
| | industrial | rqkmeans | 5 | 128 | 3105 | 659 (21.2%) | yes | |
| | industrial | rqkmeans | 5 | 256 | 3105 | 462 (14.9%) | yes | |
| | industrial | rqkmeans | 5 | 512 | 3105 | 196 (6.3%) | yes | |
| | industrial | rqvae | 3 | 128 | 3105 | 22 (0.7%) | - | |
| | industrial | rqvae | 3 | 256 | 3105 | 24 (0.8%) | yes | |
| | industrial | rqvae | 3 | 512 | 3105 | 22 (0.7%) | - | |
| | industrial | rqvae | 4 | 128 | 3105 | 26 (0.8%) | - | |
| | industrial | rqvae | 4 | 256 | 3105 | 22 (0.7%) | yes | |
| | industrial | rqvae | 4 | 512 | 3105 | 22 (0.7%) | - | |
| | industrial | rqvae | 5 | 128 | 3105 | 22 (0.7%) | - | |
| | industrial | rqvae | 5 | 256 | 3105 | 22 (0.7%) | yes | |
| | industrial | rqvae | 5 | 512 | 3105 | 79 (2.5%) | - | |
| | office | MQ | 3 | 128 | 17696 | 12324 (69.6%) | - | |
| | office | MQ | 3 | 256 | 17696 | 9713 (54.9%) | yes | |
| | office | MQ | 3 | 512 | 17696 | 7279 (41.1%) | yes | |
| | office | MQ | 4 | 128 | 17696 | 8757 (49.5%) | - | |
| | office | MQ | 4 | 256 | 17696 | 6009 (34.0%) | yes | |
| | office | MQ | 4 | 512 | 17696 | 4193 (23.7%) | yes | |
| | office | MQ | 5 | 128 | 17696 | 6174 (34.9%) | - | |
| | office | MQ | 5 | 256 | 17696 | 4185 (23.6%) | yes | |
| | office | MQ | 5 | 512 | 17696 | 2769 (15.6%) | yes | |
| | office | rqkmeans | 3 | 128 | 17696 | 10210 (57.7%) | yes | |
| | office | rqkmeans | 3 | 256 | 17696 | 7231 (40.9%) | yes | |
| | office | rqkmeans | 3 | 512 | 17696 | 5319 (30.1%) | yes | |
| | office | rqkmeans | 4 | 128 | 17696 | 5687 (32.1%) | yes | |
| | office | rqkmeans | 4 | 256 | 17696 | 4019 (22.7%) | yes | |
| | office | rqkmeans | 4 | 512 | 17696 | 3090 (17.5%) | yes | |
| | office | rqkmeans | 5 | 128 | 17696 | 3568 (20.2%) | yes | |
| | office | rqkmeans | 5 | 256 | 17696 | 2683 (15.2%) | yes | |
| | office | rqkmeans | 5 | 512 | 17696 | 2230 (12.6%) | yes | |
| | office | rqvae | 3 | 128 | 17696 | 242 (1.4%) | - | |
| | office | rqvae | 3 | 256 | 17696 | 173 (1.0%) | yes | |
| | office | rqvae | 3 | 512 | 17696 | 159 (0.9%) | - | |
| | office | rqvae | 4 | 128 | 17696 | 127 (0.7%) | - | |
| | office | rqvae | 4 | 256 | 17696 | 123 (0.7%) | yes | |
| | office | rqvae | 4 | 512 | 17696 | 117 (0.7%) | - | |
| | office | rqvae | 5 | 128 | 17696 | 121 (0.7%) | - | |
| | office | rqvae | 5 | 256 | 17696 | 122 (0.7%) | yes | |
| | office | rqvae | 5 | 512 | 17696 | 117 (0.7%) | - | |
|
|
