Row-standardise the store-page prefix in wscan_gal_rev.npz and ss_queries_rev_S.npy

README.md

@@ -1,55 +1,81 @@
----
-license: mit
-pretty_name: latenarray-I2CE data cache
-tags:
-- steam
-- game-reviews
-- sentence-embeddings
-- representation-learning
-size_categories:
-- 10M<n<100M
----
-
-# latenarray-I2CE data cache
-
-These are the precomputed tensors that the w9 workers of the I-CE code release
-(`contrast_experiment/w9/Pod/w9_cv_worker.py`, `w9_a100_worker.py`) read from their
-`--data-dir`. Download the whole folder, rebuild the full pool, and pass the folder as `--data-dir`:
-
-```bash
-hf download chironYu/latenarray-i2ce-data --repo-type dataset --local-dir latenarray-i2ce-data
-cd latenarray-i2ce-data && python join_full_pool.py
-```
-
-`full_pool_fp16.npy` (150.1 GB) is stored as four byte parts under 50 GB each,
-`full_pool_fp16.npy.part00` to `part03`. `join_full_pool.py` concatenates them, checks the
-result against `SHA256SUMS`, writes `full_pool_READY` (the marker the w9 notebooks wait for)
-and deletes the parts (`--keep-parts` keeps them).
-
-The packaged entry points of the code release (`steam_reviews_framework/run.py`,
-`contrast_experiment/run.py`) need neither the full pool nor the two `wscan_gal_rev_g*.npz`
-galleries, about 21 GiB in all; they rebuild the anchor packs at their budget from
-`wscan_pool_rev.npy` and `full_pool_meta.npz`:
-
-```bash
-hf download chironYu/latenarray-i2ce-data --repo-type dataset --local-dir data/assets \
-  --exclude "full_pool_fp16.npy.part*" "wscan_gal_rev_g*.npz"
-```
-
-The corpus covers 2,020 Steam games released 2020 to 2024: 6,649,325 reviews and
-73,289,517 sentences. Every sentence is embedded once with Qwen/Qwen3-Embedding-0.6B
-(1,024-d, last-token pooling, no instruction prefix, not normalized) and stored as float16.
-The repository holds no review text, only vectors, offsets and identifiers.
-
-| File | Contents |
-|---|---|
-| `full_pool_fp16.npy` | All 73,289,517 sentence vectors, `[73289517, 1024]` float16 (139.8 GiB). Training views are drawn from it (`--full-pool`); the 4,096-sentence evaluation draws read it as well |
-| `full_pool_meta.npz` | `game_names` (2,020), `game_review_offsets` (2,021), `review_offsets` (6,649,326): game g owns reviews `game_review_offsets[g]:[g+1]`, review r owns sentence rows `review_offsets[r]:[r+1]` |
-| `wscan_pool_rev.npy` | The 2,048-sentence review pool per game, `[2020, 2048, 1024]` float16; anchor packs above 512 sentences are rebuilt from it |
-| `wscan_pool_rev_len.npy`, `wscan_pool_rev_rid.npy` | Pool lengths and review ids |
-| `wscan_gal_rev.npz`, `wscan_gal_rev_g1024.npz`, `wscan_gal_rev_g4096.npz` | Prebuilt anchor galleries at 512, 1,024 and 4,096 sentences (the 4,096 one is a single draw used by the fixed-split worker) |
-| `games.npz` | Game names and per-game assets |
-| `wiki_eval.npz`, `wiki_eval_split.json` | The 814-game wiki evaluation universe and its split (seed 20260711) |
-| `wiki_clean_views.npz`, `wiki_llm_views.npz`, `sp_raw_views.npz` | Document views: cleaned Wikipedia, LLM rewrites, store pages |
-| `ss_queries_rev.npz`, `ss_queries_rev_S.npy` | Evaluation queries and their sentence vectors |
-| `tag_labels.npz`, `_tag_splitM.json` | Tag labels (23 coarse classes) and the tag-probe split |
+---
+license: mit
+pretty_name: latenarray-I2CE data cache
+tags:
+- steam
+- game-reviews
+- sentence-embeddings
+- representation-learning
+size_categories:
+- 10M<n<100M
+---
+
+# latenarray-I2CE data cache
+
+These are the precomputed tensors that the w9 workers of the I-CE code release
+(`contrast_experiment/w9/Pod/w9_cv_worker.py`, `w9_a100_worker.py`) read from their
+`--data-dir`. Download the whole folder, rebuild the full pool, and pass the folder as `--data-dir`:
+
+```bash
+hf download chironYu/latenarray-i2ce-data --repo-type dataset --local-dir latenarray-i2ce-data
+cd latenarray-i2ce-data && python join_full_pool.py
+```
+
+`full_pool_fp16.npy` (150.1 GB) is stored as four byte parts under 50 GB each,
+`full_pool_fp16.npy.part00` to `part03`. `join_full_pool.py` concatenates them, checks the
+result against `SHA256SUMS`, writes `full_pool_READY` (the marker the w9 notebooks wait for)
+and deletes the parts (`--keep-parts` keeps them).
+
+The packaged entry points of the code release (`steam_reviews_framework/run.py`,
+`contrast_experiment/run.py`) need neither the full pool nor the two `wscan_gal_rev_g*.npz`
+galleries, about 21 GiB in all; they rebuild the anchor packs at their budget from
+`wscan_pool_rev.npy` and `full_pool_meta.npz`:
+
+```bash
+hf download chironYu/latenarray-i2ce-data --repo-type dataset --local-dir data/assets \
+  --exclude "full_pool_fp16.npy.part*" "wscan_gal_rev_g*.npz" "*_rawprefix.*"
+```
+
+The corpus covers 2,020 Steam games released 2020 to 2024: 6,649,325 reviews and
+73,289,517 sentences. Every sentence is embedded once with Qwen/Qwen3-Embedding-0.6B
+(1,024-d, last-token pooling, no instruction prefix, not normalized) and stored as float16.
+The repository holds no review text, only vectors, offsets and identifiers.
+
+Sentence vectors are stored raw in the full pool, the document views and the wiki queries; the
+workers row-standardise them (mean 0, std 1 over the 1,024 dimensions) when they load them. The
+review pool, the anchor packs and the pseudo-queries are stored already row-standardised.
+
+| File | Contents |
+|---|---|
+| `full_pool_fp16.npy` | All 73,289,517 sentence vectors, `[73289517, 1024]` float16 (139.8 GiB). Training views are drawn from it (`--full-pool`); the 4,096-sentence evaluation draws read it as well |
+| `full_pool_meta.npz` | `game_names` (2,020), `game_review_offsets` (2,021), `review_offsets` (6,649,326): game g owns reviews `game_review_offsets[g]:[g+1]`, review r owns sentence rows `review_offsets[r]:[r+1]` |
+| `wscan_pool_rev.npy` | The 2,048-sentence review pool per game, `[2020, 2048, 1024]` float16: the 3 longest reviews of the game first (when they fit), then whole reviews at random; anchor packs above 512 sentences are rebuilt from it |
+| `wscan_pool_rev_len.npy`, `wscan_pool_rev_rid.npy` | Pool lengths and review ids |
+| `wscan_gal_rev.npz`, `wscan_gal_rev_g1024.npz`, `wscan_gal_rev_g4096.npz` | Prebuilt anchor galleries at 512, 1,024 and 4,096 sentences (the 4,096 one is a single draw used by the fixed-split worker): store-page prefix (`gal_doc_len` rows), then whole reviews; `gal_len` rows used. See *Changes* for the 512 one |
+| `wscan_gal_rev_rawprefix.npz` | The 512-sentence galleries exactly as the published @512 runs trained on them, store-page prefix not row-standardised. Only needed to reproduce those runs bit for bit |
+| `games.npz` | Game names (`names`, the game index of every other file) and per-game assets |
+| `wiki_eval.npz`, `wiki_eval_split.json` | The evaluation queries: LLM rewrites of the wiki articles of the 814-game evaluation universe, four variants each (`neutral`, `noname`, `positive`, `negative`), and the split (seed 20260711) |
+| `wiki_clean_views.npz`, `wiki_llm_views.npz`, `sp_raw_views.npz` | Document views: cleaned Wikipedia, LLM rewrites, store pages |
+| `ss_queries_rev.npz`, `ss_queries_rev_S.npy` | Pseudo-queries: four anchor-shaped packs per game (store-page prefix, then whole pool reviews to 512 sentences), flat-stored, query j = `S[off[j]:off[j+1]]`. Report-only diagnostics; checkpoint selection and every published number use `wiki_eval.npz` instead |
+| `ss_queries_rev_S_rawprefix.npy` | The pseudo-query vectors before the prefix fix (see *Changes*) |
+| `tag_labels.npz`, `_tag_splitM.json` | Tag labels (23 coarse classes) and the tag-probe split |
+
+## Changes
+
+**2026-11-07: store-page prefix row-standardised in `wscan_gal_rev.npz` and `ss_queries_rev_S.npy`.**
+Both files put a game's `sp_raw` store-page sentences in front of its review sentences. The
+review sentences were row-standardised; the store-page sentences were copied raw (row std about
+2.9 against 1.0), and the tower has no input normalisation. Every other input already reached it
+standardised, including the store-page prefix of the 1,024- and 4,096-sentence galleries and of
+the galleries the workers build on the GPU at 1,024 and 2,048 sentences. The two files now carry
+the prefix row-standardised exactly as the workers standardise `sp_raw_views.npz` when they load
+it; every review row, every length and the order are unchanged, bit for bit. The files as they
+were are kept as `wscan_gal_rev_rawprefix.npz` and `ss_queries_rev_S_rawprefix.npy`: the
+published @512 runs (and the fixed-split rows trained with 512-sentence anchors) were trained on
+the raw prefix, so to reproduce them bit for bit, put the `_rawprefix` file in place of the
+current one. All other files are unchanged.
+
+| File | SHA-256 before | SHA-256 now |
+|---|---|---|
+| `wscan_gal_rev.npz` | `e4ee5da46c5d7240065f02a3948c51c78ca9fe22d3145613c752589b4b8f789e` (now `wscan_gal_rev_rawprefix.npz`) | `7a304e875717a3cb9abffb554f460d75f20e8fea9d41b4cb1020f7096cafbe5b` |
+| `ss_queries_rev_S.npy` | `2762f528c70173f2613aad397d14f2f3abf3e1ca453d5441551ac911abf37dd4` (now `ss_queries_rev_S_rawprefix.npy`) | `413ebc66c15568ede34ecf2d7f1886a3420252e5533dfbb27c2db13efe247f67` |

ss_queries_rev_S.npy

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2762f528c70173f2613aad397d14f2f3abf3e1ca453d5441551ac911abf37dd4
+oid sha256:413ebc66c15568ede34ecf2d7f1886a3420252e5533dfbb27c2db13efe247f67
 size 8461095040

wscan_gal_rev.npz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e4ee5da46c5d7240065f02a3948c51c78ca9fe22d3145613c752589b4b8f789e
-size 2118140440
+oid sha256:7a304e875717a3cb9abffb554f460d75f20e8fea9d41b4cb1020f7096cafbe5b
+size 2118141012