Add files using upload-large-folder tool

.gitattributes

@@ -58,3 +58,49 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+tmall/tmall_session.tsv filter=lfs diff=lfs merge=lfs -text
+tmall/tmall_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+amazon_m2/amazon_m2_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+amazon_m2/amazon_m2_session.tsv filter=lfs diff=lfs merge=lfs -text
+movielens_10m/movielens_10m_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+movielens_10m/movielens_10m_session.tsv filter=lfs diff=lfs merge=lfs -text
+lastfm_360k/lastfm_360k_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+amazon_full/amazon_full_session.tsv filter=lfs diff=lfs merge=lfs -text
+bookcrossing/bookcrossing_session.tsv filter=lfs diff=lfs merge=lfs -text
+yelp/yelp_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+yelp/yelp_session.tsv filter=lfs diff=lfs merge=lfs -text
+anime/anime_session.tsv filter=lfs diff=lfs merge=lfs -text
+anime/anime_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+alibaba_ifashion/ifashion_outfit_session_v1.tsv filter=lfs diff=lfs merge=lfs -text
+alibaba_ifashion/ifashion_outfit_session_v1_remapped.txt filter=lfs diff=lfs merge=lfs -text
+alibaba_ifashion/ifashion_user_session.hg filter=lfs diff=lfs merge=lfs -text
+alibaba_ifashion/ifashion_user_session.tsv filter=lfs diff=lfs merge=lfs -text
+alibaba_ifashion/ifashion_user_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+jd/jd_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+jd/jd_session.tsv filter=lfs diff=lfs merge=lfs -text
+lastfm_1k/lastfm_1k_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+movielens_25m/movielens_25m_session.tsv filter=lfs diff=lfs merge=lfs -text
+gowalla/gowalla_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+lastfm_1k/lastfm_1k_session_mapping.txt filter=lfs diff=lfs merge=lfs -text
+lastfm_1k/lastfm_1k_session.tsv filter=lfs diff=lfs merge=lfs -text
+gowalla/gowalla_item_mapping.tsv filter=lfs diff=lfs merge=lfs -text
+gowalla/gowalla_session.tsv filter=lfs diff=lfs merge=lfs -text
+movielens_25m/movielens_25m_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+yoochoose/yoochoose_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+million_song/million_song_session.tsv filter=lfs diff=lfs merge=lfs -text
+lastfm_360k/lastfm_360k_session.tsv filter=lfs diff=lfs merge=lfs -text
+yoochoose/yoochoose_session.tsv filter=lfs diff=lfs merge=lfs -text
+mind/mind_session.tsv filter=lfs diff=lfs merge=lfs -text
+mind/mind_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+criteo_terabyte/maxembed_original/criteo_terabyte_remap.txt filter=lfs diff=lfs merge=lfs -text
+movielens_32m/movielens_32m_session.tsv filter=lfs diff=lfs merge=lfs -text
+tmall/cylon_remaps/sequential_remap.json filter=lfs diff=lfs merge=lfs -text
+tmall/cylon_remaps/seedexpand_remap.json filter=lfs diff=lfs merge=lfs -text
+movielens_32m/movielens_32m_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+tmall/cylon_remaps/maxembed_remap.json filter=lfs diff=lfs merge=lfs -text
+netflix/netflix_session.tsv filter=lfs diff=lfs merge=lfs -text
+netflix/netflix_session_remapped.txt filter=lfs diff=lfs merge=lfs -text
+criteo_terabyte/merci_filtered/criteo_terabyte_train_filtered.txt filter=lfs diff=lfs merge=lfs -text
+amazon_Home_and_Kitchen/merci_filtered/amazon_Home_and_Kitchen_test_filtered.txt filter=lfs diff=lfs merge=lfs -text
+amazon_Home_and_Kitchen/merci_filtered/amazon_Home_and_Kitchen_train_filtered.txt filter=lfs diff=lfs merge=lfs -text
+amazon_m2/cylon_remaps/sequential_remap.json filter=lfs diff=lfs merge=lfs -text

alibaba_ifashion/ifashion_outfit_session_v1.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b935aa99b6af583ea4fab1e0a7b93e65908a7f07ae238b38f62c1c1a2309f768
+size 117937038

alibaba_ifashion/ifashion_outfit_session_v1_remapped.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:271c46fc85ce03739eab802ec899da624e5f1d0b270fb1fad8c9bd66e790e30c
+size 24184832

alibaba_ifashion/ifashion_user_session.hg

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a3c8358e40cdbd2da64267ae2e0b5438c45b59e2d7498ca2c2845bfacb93e36e
+size 179159066

alibaba_ifashion/ifashion_user_session.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:73a82ff8ac765dcdb61b586c04e365cf5e47940f4cb2de8f8fbdebd8a41b47a7
+size 789923298

alibaba_ifashion/ifashion_user_session_remapped.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:abd06b07cdc1a54dc9f7290fa4f3832a018150424a03fb86a932b754f205eddc
+size 179159040

amazon_Home_and_Kitchen/maxembed_original/amazon_Home_and_Kitchen_partition.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:185f7f393253d68e4680847f716e44b08f446b756e326375dacb3a8661595a29
+size 2772112

amazon_Home_and_Kitchen/maxembed_paper_scale/amazon_Home_and_Kitchen.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc3c0c0ad22292f5cf2436250c522630849fbe93c3e00b0756e270fe990bbead
+size 38447204

amazon_Home_and_Kitchen/maxembed_paper_scale/amazon_Home_and_Kitchen_mapping.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0f8c270442030e52867b6c96eb1b433ad362ab3df8b5dae81b9f71edfcd52de
+size 2772112

amazon_Home_and_Kitchen/maxembed_paper_scale/amazon_Home_and_Kitchen_replicated_r0.1.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7b3f608efade143199b819fdcdee09b60b826b51c36ccf3be5fd083b24fee56
+size 2864528

amazon_Home_and_Kitchen/maxembed_paper_scale/amazon_Home_and_Kitchen_replicated_r0.2.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f233342c0fcc16c6fd5ebcf2844faa237de6514e1299d82da2aaa8a5d3c67ab
+size 2956944

amazon_Home_and_Kitchen/maxembed_paper_scale/amazon_Home_and_Kitchen_replicated_r0.4.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a8c0e6e0cb8bca2eb4299e54f8e6b636d583ce70eb522da4186cb5bfd3888b6
+size 3141776

amazon_Home_and_Kitchen/maxembed_paper_scale/amazon_Home_and_Kitchen_replicated_r0.8.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1f36286feb791704206f8a0f07093b2165c206d5acc09156dd81ace0ba27dc2
+size 3511376

amazon_Home_and_Kitchen/maxembed_paper_scale/amazon_Home_and_Kitchen_replicated_r0.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0f8c270442030e52867b6c96eb1b433ad362ab3df8b5dae81b9f71edfcd52de
+size 2772112

amazon_Home_and_Kitchen/merci_filtered/amazon_Home_and_Kitchen_test_filtered.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec7e75446e111feb6f0f0251e7c896075b93b53b77221bc28198d475f573d687
+size 11170482

amazon_Home_and_Kitchen/merci_filtered/amazon_Home_and_Kitchen_train_filtered.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e76b62698d76c609f41fc9e17fe392a0353ef4f75fd7784833f48b1c8b05585d
+size 12568206

amazon_Office_Products/maxembed_original/amazon_Office_Products_partition.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19e837e8e980355a2c6f6e3fd7f3178538ec753deac845e376e7e1c24bac0d20
+size 367384

amazon_Office_Products/maxembed_paper_scale/amazon_Office_Products.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b31989e748e51ea246f7c870438e05f126f6323e0153ca2f5126fc2107e8c0fa
+size 2759708

amazon_Office_Products/maxembed_paper_scale/amazon_Office_Products_mapping.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d8357607d7cc2731100b75b09e87ad97be08d7f3091873641ab363944f54c644
+size 367384

amazon_Office_Products/maxembed_paper_scale/amazon_Office_Products_replicated_r0.1.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:744af9e09ae625056810431164545b9334a8ac84c92b5820ccbebfd182b724bf
+size 379672

amazon_Office_Products/maxembed_paper_scale/amazon_Office_Products_replicated_r0.2.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14baac785a24332b84d8a70778e1c86df1ac16eab0aad5fb941d09dcf81db1f9
+size 391896

amazon_Office_Products/maxembed_paper_scale/amazon_Office_Products_replicated_r0.4.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8196f21b1cb2afb6b6c561a7a8c9ebb99790543acf43d8183775e861fecca176
+size 416408

amazon_Office_Products/maxembed_paper_scale/amazon_Office_Products_replicated_r0.8.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13b4c6eedac13bc97a1d9f4541c3c8e3e135db5c4cf5d26f31f6967ead836685
+size 465368

amazon_Office_Products/maxembed_paper_scale/amazon_Office_Products_replicated_r0.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d8357607d7cc2731100b75b09e87ad97be08d7f3091873641ab363944f54c644
+size 367384

amazon_full/amazon_full_session.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:775b75eeee24bf411624e6d753d98afea8a4dc3fc720eabb4632f2b340742bc1
+size 18581089

amazon_m2/amazon_m2_session.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3b6f01978e6c916b4b9f4526f12de72d0ec0cc8c28b389774f13464268cf328
+size 208036773

amazon_m2/amazon_m2_session_remapped.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1ae13e42f86219a9bdee0f8e6555d12ab2f9d27547f12fee584821b7d554d88
+size 136225011

amazon_m2/cylon_remaps/sequential_remap.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:168c2b845df2e71a886e3549a00dc1aa9f9d70342a74bcbdf005b8259ce68785
+size 11434607

anime/anime_session.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e288dad0e28af60862beeb192c32fa5ab3a44ec0e5f02209d3f9893ba2a4024
+size 39786210

anime/anime_session_remapped.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d4ef84c6168e27e1b02dca53935165ba222d75673b0d41deb9fcfb76a84ce32
+size 39244059

avazu/maxembed_paper_scale/avazu_paper_scale.py

@@ -0,0 +1,284 @@
+#!/usr/bin/env python3
+"""
+Paper-scale Avazu hypergraph generator.
+
+Follows MaxEmbed's official preprocess.py logic but uses efficient pandas primitives
+(pd.factorize + vectorized value_counts) to scale to the full 40.4M-row Avazu dataset.
+
+Steps replicated from MaxEmbed/partition/preprocess/avazu_script/preprocess.py:
+  1. Drop 'id' column, keep everything else as strings/ints.
+  2. Fill missing categorical with '80000000', missing integer with INT_MIN.
+  3. For categorical cols: ordinal encode by descending frequency, +1 (min=1).
+  4. For dense col (I1=hour): LabelEncoder, +1.
+  5. Apply offsets: for categorical cols, df[col] += offset, offset += col.max().
+  6. Drop first `drop_cnt` columns (label, I1 by default), leaving C1..C21.
+  7. Per-row dedup (MaxEmbed feature: overlapping offsets make some rows short).
+  8. Emit MaxEmbed hypergraph text format.
+
+Default: drop_cnt=2, normalize_dense=0, feature_cross=0.
+Input: train.csv with header row.
+Output: <prefix>_hypergraph.txt
+"""
+
+import argparse
+import logging
+import numpy as np
+import pandas as pd
+import sklearn.preprocessing as skp
+import time
+import os
+
+logging.basicConfig(format='%(asctime)s %(message)s', level=logging.INFO)
+log = logging.info
+
+NUM_INTEGER_COLUMNS = 1
+NUM_CATEGORICAL_COLUMNS = 21
+# Column names aligned with MaxEmbed script:
+#   label = click, I1 = hour, C1..C21 = remaining 21 categorical columns
+# Source CSV cols (post-drop-id): click, hour, C1, banner_pos, site_id, site_domain,
+#   site_category, app_id, app_domain, app_category, device_id, device_ip, device_model,
+#   device_type, device_conn_type, C14, C15, C16, C17, C18, C19, C20, C21
+SCRIPT_COLS = ['label', 'I1'] + [f'C{i}' for i in range(1, NUM_CATEGORICAL_COLUMNS + 1)]
+CAT_COLS = SCRIPT_COLS[2:]   # C1..C21
+DENSE_COLS = ['I1']
+LABEL_COL = 'label'
+CAT_NAN_VALUE = '80000000'
+INT_NAN_VALUE = np.iinfo(np.int32).min
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument('--src_csv_path', required=True)
+    ap.add_argument('--output_prefix', required=True,
+                    help='Output prefix. Writes <prefix>_hypergraph.txt (+ optional .bin).')
+    ap.add_argument('--drop_cnt', type=int, default=2,
+                    help='Leading columns to drop (0=keep label+I1+C*, 2=drop label+I1).')
+    ap.add_argument('--normalize_dense', type=int, default=0,
+                    help='If 1, MinMax-scale I1 to [0,1]. Default 0 so values stay integer.')
+    ap.add_argument('--binary', type=int, default=1,
+                    help='Also write binary trans_query format .bin file.')
+    ap.add_argument('--dedup_per_row', type=int, default=1,
+                    help='Deduplicate overlapping vertex IDs within each query (MaxEmbed style).')
+    args = ap.parse_args()
+
+    t0 = time.time()
+    log(f'Reading CSV from {args.src_csv_path}')
+    # Read full CSV, skip header, assign SCRIPT_COLS names.
+    # First column in source is 'id' which we skip via usecols.
+    src_cols_total = 1 + 1 + 1 + NUM_CATEGORICAL_COLUMNS   # id + label + I1 + 21 cats
+    usecols = list(range(1, src_cols_total))   # skip id (col 0)
+
+    # Use category-aware dtypes during read for memory:
+    # label + I1 = int32 (after fillna), cats = string (factorize later).
+    dtype_map = {}
+    for i, c in enumerate(SCRIPT_COLS):
+        dtype_map[c] = 'string'   # read as string first, cast later
+
+    df = pd.read_csv(args.src_csv_path,
+                     header=0,
+                     names=SCRIPT_COLS,
+                     usecols=usecols,
+                     dtype=dtype_map,
+                     engine='c',
+                     low_memory=True)
+    log(f'CSV loaded: shape={df.shape} in {time.time()-t0:.1f}s')
+
+    # ---- Fill missing ----
+    log('Filling missing values')
+    for c in CAT_COLS:
+        df[c] = df[c].fillna(CAT_NAN_VALUE)
+    # label and I1 are numeric. Fill with INT_NAN then cast.
+    df[LABEL_COL] = df[LABEL_COL].fillna(str(INT_NAN_VALUE)).astype(np.int64)
+    df['I1'] = df['I1'].fillna(str(INT_NAN_VALUE)).astype(np.int64)
+
+    # ---- Transform per column ----
+    # Label and dense: LabelEncoder (fit+transform). Same as script: codes + 1 for dense (not label).
+    log('Encoding label via LabelEncoder')
+    le = skp.LabelEncoder()
+    df[LABEL_COL] = le.fit_transform(df[LABEL_COL].values).astype(np.int64)
+
+    log('Encoding I1 via LabelEncoder, +1')
+    le = skp.LabelEncoder()
+    df['I1'] = le.fit_transform(df['I1'].values).astype(np.int64) + 1
+
+    # We apply `+1` to match the official script so that 0 can be reserved, but
+    # later we must produce 0-indexed IDs to satisfy MaxEmbed's `num < c` invariant.
+    # We compensate by subtracting 1 at the final write stage.
+
+    if args.normalize_dense:
+        mms = skp.MinMaxScaler(feature_range=(0, 1))
+        df['I1'] = mms.fit_transform(df[['I1']].values).ravel()
+        # Note: normalizing integer IDs is weird; we leave this for compat.
+
+    # Categorical: ordinal encode by descending frequency, +1
+    # Original script uses value_counts() and applies .index.get_loc() per row (O(N*K)).
+    # Equivalent & fast: pd.factorize then remap codes by frequency rank.
+    for c in CAT_COLS:
+        tt = time.time()
+        codes, uniques = pd.factorize(df[c], sort=False)
+        # Frequency of each code
+        counts = np.bincount(codes.astype(np.int64, copy=False))
+        # Order descending by count. argsort on descending gives the new rank.
+        # rank[old_code] = new_code (0..K-1 by desc freq)
+        order = np.argsort(-counts, kind='stable')
+        new_rank = np.empty_like(order)
+        new_rank[order] = np.arange(len(order))
+        df[c] = (new_rank[codes] + 1).astype(np.int64)   # +1 so min=1
+        log(f'  col {c}: unique={len(uniques):>10,}  in {time.time()-tt:.1f}s')
+
+    # ---- Min/max per column ----
+    # Matches original script's min_max collection (used for offset bookkeeping).
+    min_max = {}
+    for c in SCRIPT_COLS:
+        if c == LABEL_COL or c.startswith('I'):
+            mn = int(df[c].min()); mx = int(df[c].max())
+            min_max[c] = (mn, mx)
+        else:
+            mn = int(df[c].min()); mx = int(df[c].max())
+            min_max[c] = (mn, mx)
+        log(f'  min_max[{c}] = [{min_max[c][0]}, {min_max[c][1]}]')
+
+    # ---- Apply offsets to categorical columns ----
+    log('Applying column offsets')
+    offset = np.int64(0)
+    for c in SCRIPT_COLS:
+        if c == LABEL_COL or c.startswith('I'):
+            continue
+        df[c] += offset
+        log(f'  {c} offset={offset}')
+        offset += min_max[c][1]
+
+    total_vocab = int(offset)
+    log(f'Total offset (vocab across cat cols) = {total_vocab}')
+
+    # ---- Drop leading columns ----
+    cols_after_drop = SCRIPT_COLS[args.drop_cnt:]
+    log(f'Keeping columns: {cols_after_drop}')
+
+    # ---- Write hypergraph ----
+    out_txt = args.output_prefix + '_hypergraph.txt'
+    out_bin = args.output_prefix + '.bin'
+    log(f'Writing {out_txt}')
+
+    # Convert selected columns to int32 numpy for fast serialization
+    # (I1 can be float if normalize_dense=1; we coerce to int to fit text format unless normalized.)
+    kept_arrays = []
+    for c in cols_after_drop:
+        if c == 'I1' and args.normalize_dense:
+            kept_arrays.append(df[c].values.astype(np.float64))
+        else:
+            kept_arrays.append(df[c].values.astype(np.int64))
+    mat = np.column_stack(kept_arrays)
+    num_queries = mat.shape[0]
+    ncols = mat.shape[1]
+    log(f'Row-matrix built: {mat.shape}')
+
+    # Compute num_vertices and total_pins.
+    # vertices = unique global IDs across kept columns (excluding label which is a different "type"
+    # but original script uses drop_cnt=2 so label is dropped; if drop_cnt=0 we include label too).
+    # Since IDs are already offset, unique across mat.flatten() gives exact vocab.
+    # For efficiency with 800M+ entries, use np.unique (uses radix sort in numpy 2.x and is OK here).
+    log('Computing unique vertices (this may take 1-2 min)')
+    t1 = time.time()
+    all_vals = mat.ravel()
+    if args.dedup_per_row:
+        # Per-row dedup for MaxEmbed format: sort within row and drop duplicates.
+        # We implement this while writing to avoid huge intermediate.
+        # For vertex count: just use unique over all.
+        uniq_vertices = np.unique(all_vals)
+        num_vertices = uniq_vertices.size
+        log(f'  unique vertices = {num_vertices:,} (in {time.time()-t1:.1f}s)')
+
+        # Write file streaming, with per-row dedup.
+        # Pass 1: compute per-row dedup'd rows into a list-of-arrays, also sum pins.
+        # Then write file with correct header in one pass. At 40M queries with avg 20 pins,
+        # storing arrays is ~3.2GB (uint32) which fits in 197GB RAM.
+        log('Pass 1: per-row sort + dedup + pin count')
+        t1 = time.time()
+        # We'll precompute CSR (offsets + all_data) which is also what .bin needs.
+        # Upper bound on total pins = num_queries * ncols.
+        all_data = np.empty(num_queries * ncols, dtype=np.uint32)
+        offsets_arr = np.zeros(num_queries + 1, dtype=np.uint64)
+        cursor = 0
+        CHUNK = 262144
+        for start in range(0, num_queries, CHUNK):
+            end = min(start + CHUNK, num_queries)
+            block = mat[start:end]
+            block_sorted = np.sort(block, axis=1)
+            diff = np.diff(block_sorted, axis=1, prepend=np.int64(-1))
+            keep = diff != 0
+            # Per-row write into all_data
+            for i in range(block_sorted.shape[0]):
+                row = block_sorted[i][keep[i]]
+                all_data[cursor:cursor+row.size] = row.astype(np.uint32, copy=False)
+                cursor += row.size
+                offsets_arr[start + i + 1] = cursor
+            if start % (CHUNK * 40) == 0:
+                log(f'  rows processed {end:,}/{num_queries:,}  pins={cursor:,}  '
+                    f'elapsed={time.time()-t1:.0f}s')
+        total_pins = cursor
+        # Truncate all_data
+        all_data = all_data[:total_pins]
+        log(f'Pass 1 done: total_pins={total_pins:,} in {time.time()-t1:.0f}s')
+
+        log(f'Pass 2: writing {out_txt}')
+        t1 = time.time()
+        with open(out_txt, 'w') as f:
+            f.write('0 {} {} {}\n'.format(num_vertices, num_queries, total_pins))
+            # Write rows. Use numpy savetxt-like approach per chunk for speed.
+            for start in range(0, num_queries, CHUNK):
+                end = min(start + CHUNK, num_queries)
+                chunk_off_start = offsets_arr[start]
+                chunk_off_end = offsets_arr[end]
+                # Build text per row
+                lines = []
+                for i in range(start, end):
+                    r = all_data[offsets_arr[i]:offsets_arr[i+1]]
+                    lines.append(' '.join(map(str, r.tolist())))
+                f.write('\n'.join(lines))
+                f.write('\n')
+                if start % (CHUNK * 40) == 0:
+                    log(f'  rows written {end:,}/{num_queries:,}  elapsed={time.time()-t1:.0f}s')
+        log(f'  wrote {out_txt}  in {time.time()-t1:.0f}s')
+    else:
+        # No dedup: all rows have fixed width = ncols
+        total_pins = num_queries * ncols
+        uniq_vertices = np.unique(all_vals)
+        num_vertices = uniq_vertices.size
+        log(f'  unique vertices = {num_vertices:,}  total_pins = {total_pins:,}')
+        log('Writing hypergraph (no dedup)')
+        with open(out_txt, 'w') as f:
+            f.write('0 {} {} {}\n'.format(num_vertices, num_queries, total_pins))
+            CHUNK = 262144
+            for start in range(0, num_queries, CHUNK):
+                end = min(start + CHUNK, num_queries)
+                block = mat[start:end]
+                lines = ['\n'.join(' '.join(map(str, row.tolist())) for row in block)]
+                f.write(lines[0] + '\n')
+
+    log(f'Vocab summary: num_vertices={num_vertices:,}  num_queries={num_queries:,}  total_pins={total_pins:,}')
+    sz = os.path.getsize(out_txt)
+    log(f'{out_txt}: {sz/1e6:.1f} MB')
+
+    # Optional: write binary .bin format matching trans_query.cpp
+    if args.binary and args.dedup_per_row:
+        log(f'Writing binary {out_bin}')
+        # Format from trans_query.cpp:
+        #   uint64 query_cnt
+        #   uint64 node_cnt
+        #   uint64 index[query_cnt+1]  (CSR offsets into all_data)
+        #   uint32 all_data[total_pins]
+        with open(out_bin, 'wb') as f:
+            f.write(np.uint64(num_queries).tobytes())
+            f.write(np.uint64(num_vertices).tobytes())
+            f.write(offsets_arr.tobytes())
+            f.write(all_data.tobytes())
+        log(f'  wrote {out_bin}  size={os.path.getsize(out_bin)/1e6:.1f} MB')
+    elif args.binary:
+        log('--binary requires --dedup_per_row=1 (no-dedup binary path not implemented)')
+
+    log(f'Done in {time.time()-t0:.0f}s')
+
+
+if __name__ == '__main__':
+    main()

avazu/maxembed_paper_scale/shift_ids.py

@@ -0,0 +1,64 @@
+#!/usr/bin/env python3
+"""Shift all vertex IDs by -1 in-place for the avazu hypergraph to match
+MaxEmbed's `num < c` invariant (max must be < num_vertices).
+
+Applied to both avazu.bin (binary) and avazu_hypergraph.txt (text)."""
+import numpy as np
+import os
+import time
+import struct
+
+txt_path = '/home/sadoo/research_data/processed/avazu/maxembed_paper_scale/avazu_hypergraph.txt'
+bin_path = '/home/sadoo/research_data/processed/avazu/maxembed_paper_scale/avazu.bin'
+
+# --- Binary: shift pins by 1 ---
+t0 = time.time()
+print(f'Shifting binary {bin_path}')
+with open(bin_path, 'rb') as f:
+    query_cnt = struct.unpack('<Q', f.read(8))[0]
+    node_cnt = struct.unpack('<Q', f.read(8))[0]
+    print(f'  query_cnt={query_cnt:,} node_cnt={node_cnt:,}')
+    offsets = np.frombuffer(f.read((query_cnt+1) * 8), dtype=np.uint64).copy()
+    pins = np.fromfile(f, dtype=np.uint32)
+print(f'  loaded pins: {len(pins):,}  max={pins.max()}  min={pins.min()}  (elapsed {time.time()-t0:.0f}s)')
+pins -= 1
+print(f'  after shift: max={pins.max()}  min={pins.min()}')
+
+tmp_path = bin_path + '.tmp'
+with open(tmp_path, 'wb') as f:
+    f.write(np.uint64(query_cnt).tobytes())
+    f.write(np.uint64(node_cnt).tobytes())
+    f.write(offsets.tobytes())
+    pins.tofile(f)
+os.rename(tmp_path, bin_path)
+print(f'  binary rewritten in {time.time()-t0:.0f}s, size={os.path.getsize(bin_path):,}')
+
+# --- Text: rewrite with IDs shifted by 1 using numpy-backed CSR reconstruction ---
+# We already have offsets + pins in memory (shifted), just re-emit text from them.
+t0 = time.time()
+print(f'Shifting text {txt_path}')
+tmp_path = txt_path + '.tmp'
+# Header: keep same (num_vertices was already correct = 9,449,205)
+# but now vertex IDs are 0-indexed, so max=9449204 < num_vertices=9449205.
+with open(tmp_path, 'w') as fout:
+    fout.write(f'0 {int(node_cnt)} {int(query_cnt)} {len(pins)}\n')
+    CHUNK = 262144
+    # Use buffered writes via join of per-row strings
+    cursor = 0
+    for start in range(0, int(query_cnt), CHUNK):
+        end = min(start + CHUNK, int(query_cnt))
+        # per-row slicing
+        buf = []
+        for i in range(start, end):
+            lo = int(offsets[i])
+            hi = int(offsets[i+1])
+            row = pins[lo:hi]
+            buf.append(' '.join(map(str, row.tolist())))
+        fout.write('\n'.join(buf))
+        fout.write('\n')
+        if start % (CHUNK * 40) == 0:
+            print(f'  rows written {end:,}/{int(query_cnt):,}  elapsed={time.time()-t0:.0f}s')
+os.rename(tmp_path, txt_path)
+print(f'  text rewritten in {time.time()-t0:.0f}s  size={os.path.getsize(txt_path):,}')
+
+print('Done')

bookcrossing/bookcrossing_session.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d267c11e42aaa07e9ae9d74f4ff34e8cfd172456079c9f4c88ed24e038a2ac78
+size 11998288

criteo_terabyte/maxembed_original/criteo_terabyte_partition.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da1b5368e0dcd427fcce8646e196dda19eaf303e5800fc45f4291a7820d03479
+size 10573588

criteo_terabyte/maxembed_original/criteo_terabyte_remap.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3539a7544a6e849b4bb464a17009379eff0d10aadb53235c9f134c81f3a68c58
+size 11335863

criteo_terabyte/merci_filtered/criteo_terabyte_train_filtered.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8078377cc734dc4ad724f5bc405b8f47b1626e7e264a5806d20a098fccd8a7ff
+size 46032900

foursquare/maxembed_paper_scale/foursquare.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca9c1a67406a1c63b442f85d8b554178d574d57150216a5847c33f23d98dc219
+size 1230284

foursquare/maxembed_paper_scale/foursquare_mapping.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26cc695cad773898ac6bb18ab0151580ddf4c0e58420b61b63a15ce65590ef38
+size 1202308

foursquare/maxembed_paper_scale/foursquare_replicated_r0.1.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bbf6a10a23d0a0d823d2a9b3c61ccedbeca3bb107321454afd8b7dbab3a185d6
+size 1242436

foursquare/maxembed_paper_scale/foursquare_replicated_r0.2.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:890b616cb374866751d362085b8d9bee6e39e89a878f60b228437268d072a69f
+size 1282500

foursquare/maxembed_paper_scale/foursquare_replicated_r0.4.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e97a2d078c9582c30e14d447243b4933badf0341e5d79be9c974c4d778792ece
+size 1362628

foursquare/maxembed_paper_scale/foursquare_replicated_r0.8.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2373c72d543be32e49171547df5eb2fee5882734c4154530281a4eb023eaaab9
+size 1522948

foursquare/maxembed_paper_scale/foursquare_replicated_r0.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26cc695cad773898ac6bb18ab0151580ddf4c0e58420b61b63a15ce65590ef38
+size 1202308

gowalla/gowalla_item_mapping.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:21508488a836c3308bf68dc473e750cdc61d2ea24902e21f76d4a44877fdc438
+size 18570235

gowalla/gowalla_session.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a50a89c108ff51c0e4b6f6036abf0bfc07dfefb3534827d6e7be1b3505541e35
+size 27851812

gowalla/gowalla_session_remapped.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc0673c93e826d93a6cd5e1b4c9570ee3749c1651adf77fb8252f1b78a4ead78
+size 28324176