revison-1

DATASHEET.md

@@ -1,32 +1,50 @@
 # Datasheet for the ViHoRec Dataset
 
-Following the *Datasheets for Datasets* framework (Gebru et al., 2021). All
-statistics below are produced automatically by `scripts/quality_control.py`,
-`scripts/anonymize.py`, and `scripts/make_benchmark_split.py`.
+Revision 1 of the public release. Statistics match the revised manuscript:
+17,911 interactions, 6,822 reviewer keys, and a three-way temporal split
+(798 keys; 8,645 train / 798 validation / 798 test).
+
+Following the *Datasheets for Datasets* framework (Gebru et al., 2021).
+Counts are produced by `scripts/quality_control.py`, `scripts/anonymize.py`,
+and `scripts/make_benchmark_split.py` in
+[MinhNguyenDS/ViHoRec](https://github.com/MinhNguyenDS/ViHoRec).
 
 ## 1. Motivation
 - **Purpose.** There is no publicly documented Vietnamese hotel recommendation
   dataset. ViHoRec fills this gap for research on collaborative filtering,
-  content-based, and hybrid recommendation, and on cold-start handling.
-- **Created by.** The authors (University of Information Technology, VNU-HCM).
+  content-based, and hybrid recommendation, and on short-history ranking.
+  The full corpus is cold-start dominated (most reviewer keys have one
+  interaction); the public protocol is short-history, not strict cold-start.
+- **Created by.** The authors (University of Information Technology, VNU-HCM,
+  and University of Science, VNU-HCM).
 - **Not for.** Commercial use (see LICENSE) or re-identification of individuals.
 
 ## 2. Composition
-Three released tables (`release/`):
 
-| File | Rows | Columns |
+| Hub subset | Rows | Columns |
 |---|---|---|
-| `interactions.csv` | 18,267 | user_id, hotel_id, rating, date, source |
-| `users.csv` | 6,832 | user_id, n_interactions |
-| `hotels.csv` | 560 | hotel_id, name, location |
-| content metadata (`data_content_based_raw.csv`) | 309 | 11 attributes (facilities, surroundings, vicinity, price, distance, ...) |
+| `interactions` | 17,911 | user_id, hotel_id, rating, date, source |
+| `users` | 6,822 | user_id, n_interactions |
+| `hotels` | 560 | hotel_id, name, location |
+| `content` | 310 rows / 309 names | 11 attributes (see below) |
+| `benchmark` | 8,645 / 798 / 798 | userID, itemID, rating, timestamp |
+| `benchmark_ge7` | 8,645 / 506 / 512 | same schema; holdouts rated ≥ 7 |
 
-- **Instances.** A row in `interactions.csv` is one user–hotel rating (0–10)
-  with a timestamp and its originating site.
-- **Sources.** Booking.com (7,597), Traveloka (6,273), Ivivu (4,404).
-- **Ratings** span 1.0–10.0; **dates** span 2011-10-15 to 2023-12-09.
-- **Sensitive data.** Direct identifiers (reviewer display names) are **removed**
-  before release; user ids are salted-HMAC pseudonyms (see §6).
+- **Instances.** A row in `interactions` is one reviewer-key–hotel rating on
+  the [1, 10] scale (mean 7.58), with a date and its originating site.
+- **Sources (raw crawl).** Booking.com 7,597; Traveloka 6,273; Ivivu 4,404.
+  After dropping exact duplicates and placeholder names, the released rows
+  are Booking.com 7,239; Traveloka 6,273; Ivivu 4,399.
+- **Dates** span 2011-10-15 to 2023-12-09.
+- **Content metadata.** Facilities, surroundings (Around), vicinity, price,
+  plus name, location, overall rating, review count, quality, distance to
+  center, and booking link. The sheet has 310 rows and 309 distinct names:
+  Mövenpick Villas & Residences Phu Quoc appears twice, once from Booking.com
+  and once from Traveloka, with different attributes. The manuscript count
+  of 309 is distinct names. Text from this sheet covers 283 of 535 benchmark
+  hotels (52.9%).
+- **Sensitive data.** Reviewer display names are removed before release.
+  `user_id` is a salted-HMAC pseudonym (see §6), not an anonymized person id.
 
 ## 3. Collection Process
 - **How.** Automated crawling with `requests`/BeautifulSoup and JSON review
@@ -37,54 +55,71 @@ Three released tables (`release/`):
 - **Timeframe.** Reviews were posted 2011–2023; crawling performed in 2023.
 
 ## 4. Preprocessing / Cleaning / Quality Control
-Reproduced by `scripts/quality_control.py`. Reported measures:
 
 | Check | Result |
 |---|---|
-| Field completeness | 0% missing after collection-time imputation (see limitation below) |
+| Raw interactions | 18,274 |
+| Placeholder reviewer names | 357 rows (1.95%); 356 dropped (`Không tên`, `Guest`, and variants) |
+| Other fields missing | 0.0% |
 | Exact duplicate interactions | 7 (0.038%) removed |
-| Near-duplicates (reviewer + canonical hotel + date) | 11 (0.060%) |
+| Near-duplicates (reviewer + canonical hotel + date) | 11 (0.060%); 4 retained because the rating differs |
 | Invalid / out-of-range ratings | 0 (dirty token `8..5` repaired) |
 | Unparsable dates | 0 |
-| Raw hotel names → canonical hotels | 581 → 560 (21 spelling variants merged, 3.6%) |
-| Hotels appearing on ≥2 sites | 78 |
-| Hotels with conflicting location | 1 (flagged) |
+| Raw hotel names → canonical hotels | 581 → 560 |
+| Hotels appearing on ��2 sites | 81 |
+| Hotels with conflicting location | 0 |
+| Interactions after cleaning | 17,911 |
+| Reviewer keys / hotels | 6,822 / 560 |
 
-- **Entity resolution.** Cross-site hotel matching uses an accent-free,
-  stopword-stripped, order-independent canonical key (`textnorm.py`), replacing
-  the original naïve `LabelEncoder(NameHotel)` exact-string matching.
-- **Manual validation.** `annotation_agreement.py` draws a stratified sample
-  (default n≈250: interactions + hotels) for ≥2 annotators and reports percent
-  agreement and Cohen's / Fleiss' κ, plus an estimated record-accuracy rate.
+- **Entity resolution.** Listings merge when they share a city and the same
+  discriminative name tokens, and their stated property types are compatible.
+  Pair-level validation is documented in the GitHub `reports/er_validation.md`.
+- **Identity.** `user_id` is a reviewer-name key, not a person. A same-day
+  multi-city audit is a floor of 69 colliding keys (1.01%), covering 3,513
+  interactions (19.6%). The longest remaining history is 210. The previous
+  342-review bucket was a placeholder name and is not in this release.
+- **Public split.** Three-way temporal leave-one-out. A key enters when it
+  has at least four interactions. Per key, chronological order: the last
+  interaction is test, the second-last is validation, and the remainder is
+  train. The matrix is 798 keys × 535 hotels: 8,645 / 798 / 798, sparsity
+  97.60%. The shortest training history is 2 (183 keys). See
+  `benchmark/PROTOCOL.md` and `benchmark/split_config.json`.
+- **Relevance.** Implicit next reviewed hotel. Every crawled rating is a
+  positive. `benchmark_ge7` is the sensitivity fold that keeps only holdouts
+  rated at least 7 (506 validation / 512 test users). Training rows are unchanged.
+- **Manual validation.** A stratified sample of 248 records (186 interactions
+  and 62 hotels). Checklist pass rates and the hotel-panel percent agreement
+  are reported in the manuscript.
 
 ## 5. Uses
-- Recommended: benchmarking CF/CB/hybrid recommenders, cold-start studies,
-  Vietnamese-language RecSys, low-resource / sparse-data research.
+- Recommended: benchmarking CF, content-based, and hybrid recommenders;
+  short-history studies; Vietnamese-language RecSys; sparse-data research.
+  Keys with a single interaction stay in the corpus and are excluded from scoring.
 - **Known limitations.**
-  - Small scale (18k interactions) vs. MovieLens-100k / Amazon; sparse
-    (97.5% sparsity in the benchmark split).
-  - Reviewer display names were partially imputed/normalised at crawl time
-    (missing names were replaced), so `n_interactions` per user and the
-    number of distinct users are approximate — user identity is derived from
-    a low-cardinality name string and may merge distinct individuals.
-  - Ratings are aggregate scores, not multi-criteria.
+  - Small scale (17,911 interactions); the benchmark matrix is 97.60% sparse.
+  - Reviewer keys come from display names. 6,822 is an upper bound on
+    identifiers and a lower bound on people.
+  - Content text covers 52.9% of benchmark hotels.
+  - Ratings are a single aggregate score, not multi-criteria.
+  - Hyper-parameters in `benchmark/baseline_ci.md` were selected on validation.
+    Do not retune them on `test`.
 
-## 6. Ethics, Terms of Service & Legal
+## 6. Ethics, Terms of Service and Legal
 - **Terms of Service.** Booking.com, Traveloka, and Ivivu restrict automated
-  scraping and commercial reuse in their ToS. To stay within a defensible
-  research-use position we: (a) collected only publicly visible review text and
-  ratings, no private/account data; (b) do **not** redistribute raw HTML or
-  full review text, only derived numeric ratings and hotel metadata;
-  (c) release under **CC BY-NC 4.0** (non-commercial); (d) provide takedown on
-  request. Users of this dataset must comply with the source platforms' ToS.
-- **Personal data / anonymisation.** No emails, account ids, or full names are
-  released. `anonymize.py` drops the display name entirely and assigns a
-  salted `HMAC-SHA256(secret_salt, name)[:12]` pseudonym; the secret salt is
-  kept off-repo (`VIHOREC_SALT`) and the name→id lookup
-  (`reports/_private_mapping.csv`) is **never** published.
-- **Risk.** Re-identification risk is low: no free-text, no geolocation beyond
-  city, and pseudonymous ids.
+  scraping and commercial reuse. This release: (a) collected only publicly
+  visible ratings, dates, hotel names, and city locations; (b) does not
+  redistribute raw HTML or full review text; (c) uses CC BY-NC 4.0;
+  (d) accepts takedown requests. Users must still comply with the source
+  platforms' terms in their own jurisdiction.
+- **Personal data / pseudonymization.** No emails, account ids, or display
+  names are released. `anonymize.py` assigns
+  `HMAC-SHA256(secret_salt, name)[:12]`. The salt is kept off-repository
+  (`VIHOREC_SALT`). The name-to-id lookup is never published. HMAC is not
+  anonymization.
+- **Risk.** Residual re-identification risk is non-zero, mainly for prolific
+  reviewers with distinctive rating sequences. Mitigations: no review text,
+  city-level location only, unpublished name map.
 
-## 7. Distribution & Maintenance
-- Hosted with a versioned DOI (e.g., Zenodo); this repository is the canonical
-  build pipeline. Report issues / request takedown to the corresponding author.
+## 7. Distribution and Maintenance
+- This repository is the revision-1 Hub build. The construction pipeline
+  lives on GitHub. Report issues or request takedown from the corresponding author.

LICENSE

@@ -1,9 +1,9 @@
 ViHoRec Dataset License
 =======================
 
-The ViHoRec dataset (the anonymised files under `dataset_release/release/`:
-interactions.csv, users.csv, hotels.csv, and the benchmark split) is released
-under the
+The ViHoRec dataset (the pseudonymised files in this repository:
+interactions, users, hotels, content metadata, and the benchmark splits)
+is released under the
 
     Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
 
@@ -19,9 +19,13 @@ Under the following terms:
     changes were made.
   * NonCommercial - You may not use the material for commercial purposes.
 
-The accompanying source code under `dataset_release/scripts/` is released under
+The accompanying source code in the GitHub repository is released under
 the MIT License and may be reused, including commercially.
 
+User identifiers are HMAC-SHA256 pseudonyms of reviewer display names.
+That is not anonymization: a holder of the secret salt could re-identify
+a key, and common names can still collide. See DATASHEET.md.
+
 Rationale for NonCommercial
 ---------------------------
 The underlying reviews were published by users on third-party booking platforms
@@ -34,9 +38,13 @@ CITATION
 --------
 If you use this dataset, please cite:
 
-@article{vihorec,
-  title  = {ViHoRec: A Vietnamese Hotel Recommendation Dataset with a
-            Documented Collection and Quality-Control Pipeline},
-  author = {Nguyen, Hoang Minh},
-  year   = {2026}
+@misc{nguyen2026vihorec,
+  title        = {ViHoRec: A Quality-Controlled Vietnamese Hotel Recommendation
+                  Dataset and Short-History Benchmark},
+  author       = {Nguyen, Minh Hoang and Huynh, Tin Van and Nguyen, Kiet Van},
+  year         = {2026},
+  eprint       = {2607.12946},
+  archivePrefix= {arXiv},
+  primaryClass = {cs.IR},
+  url          = {https://arxiv.org/abs/2607.12946}
 }

README.md

@@ -12,7 +12,7 @@ tags:
 - recommender-systems
 - hotel
 - vietnamese
-- cold-start
+- short-history
 - tabular
 - entity-resolution
 - parquet
@@ -31,12 +31,26 @@ configs:
   data_files:
   - split: train
     path: data/hotels/train-*.parquet
+- config_name: content
+  data_files:
+  - split: train
+    path: data/content/train-*.parquet
 - config_name: benchmark
   data_files:
   - split: train
     path: data/benchmark/train-*.parquet
+  - split: validation
+    path: data/benchmark/validation-*.parquet
   - split: test
     path: data/benchmark/test-*.parquet
+- config_name: benchmark_ge7
+  data_files:
+  - split: train
+    path: data/benchmark_ge7/train-*.parquet
+  - split: validation
+    path: data/benchmark_ge7/validation-*.parquet
+  - split: test
+    path: data/benchmark_ge7/test-*.parquet
 dataset_info:
 - config_name: interactions
   features:
@@ -52,7 +66,7 @@ dataset_info:
     dtype: string
   splits:
   - name: train
-    num_examples: 18267
+    num_examples: 17911
 - config_name: users
   features:
   - name: user_id
@@ -61,7 +75,7 @@ dataset_info:
     dtype: int64
   splits:
   - name: train
-    num_examples: 6832
+    num_examples: 6822
 - config_name: hotels
   features:
   - name: hotel_id
@@ -73,6 +87,33 @@ dataset_info:
   splits:
   - name: train
     num_examples: 560
+- config_name: content
+  features:
+  - name: Location
+    dtype: string
+  - name: NameHotel
+    dtype: string
+  - name: RatingHotel
+    dtype: float64
+  - name: CountRating
+    dtype: int64
+  - name: Price
+    dtype: int64
+  - name: Facilities
+    dtype: string
+  - name: Quality
+    dtype: float64
+  - name: DistanceCenter
+    dtype: float64
+  - name: Around
+    dtype: string
+  - name: Vicinity
+    dtype: string
+  - name: Link
+    dtype: string
+  splits:
+  - name: train
+    num_examples: 310
 - config_name: benchmark
   features:
   - name: userID
@@ -85,85 +126,113 @@ dataset_info:
     dtype: int64
   splits:
   - name: train
-    num_examples: 9787
+    num_examples: 8645
+  - name: validation
+    num_examples: 798
   - name: test
-    num_examples: 800
+    num_examples: 798
+- config_name: benchmark_ge7
+  features:
+  - name: userID
+    dtype: int64
+  - name: itemID
+    dtype: int64
+  - name: rating
+    dtype: float32
+  - name: timestamp
+    dtype: int64
+  splits:
+  - name: train
+    num_examples: 8645
+  - name: validation
+    num_examples: 506
+  - name: test
+    num_examples: 512
 ---
 
 # ViHoRec — Vietnamese Hotel Recommendation Dataset
 
-Paper: [arXiv:2607.12946](https://arxiv.org/abs/2607.12946) · [Hugging Face Papers](https://huggingface.co/papers/2607.12946)
+Revision 1 public release. Paper: [arXiv:2607.12946](https://arxiv.org/abs/2607.12946) · [Hugging Face Papers](https://huggingface.co/papers/2607.12946)
 
-A **quality-controlled, anonymised, benchmark-ready** Vietnamese hotel
-recommendation dataset for recommender-systems research.
+A **quality-controlled, pseudonymised, benchmark-ready** Vietnamese hotel
+recommendation dataset. `user_id` is an HMAC of a reviewer display name
+(a reviewer key, not an anonymised person). The public protocol is a
+**short-history** three-way temporal split, not a strict cold-start benchmark.
 
 | Resource | Count |
 |---|---|
-| Interactions (cleaned) | 18,267 |
-| Users | 6,832 |
+| Interactions (cleaned) | 17,911 |
+| Reviewer keys | 6,822 |
 | Hotels | 560 |
-| Benchmark split | 800 users × 535 hotels (9,787 train / 800 test) |
+| Content sheet | 310 rows, 309 distinct names |
+| Benchmark split | 798 keys × 535 hotels (8,645 train / 798 validation / 798 test) |
 
 Sources: Booking.com, Traveloka, Ivivu. License: **CC BY-NC 4.0** (data), MIT (code on GitHub).
 
+This card replaces the previous Hub build (18,267 interactions, 6,832 keys, two-way 800 / 9,787 / 800). Placeholder reviewer names are dropped, and hyper-parameters are selected on validation.
+
 ## Dataset Viewer / subsets
 
 The Hub Dataset Viewer is configured via the YAML `configs` block above.
-Use the **Subset** dropdown:
+Use the **Subset** dropdown. The validation split is the manuscript `val.csv`.
 
 | Subset | Splits | Rows | Description |
 |---|---|---|---|
-| `interactions` (default) | `train` | 18,267 | user–hotel ratings |
-| `users` | `train` | 6,832 | user aggregates |
-| `hotels` | `train` | 560 | hotel metadata |
-| `benchmark` | `train` / `test` | 9,787 / 800 | public LOO split |
+| `interactions` (default) | `train` | 17,911 | reviewer-key–hotel ratings |
+| `users` | `train` | 6,822 | key aggregates |
+| `hotels` | `train` | 560 | canonical hotel catalogue |
+| `content` | `train` | 310 | hotel text metadata (11 attributes) |
+| `benchmark` | `train` / `validation` / `test` | 8,645 / 798 / 798 | public three-way split |
+| `benchmark_ge7` | `train` / `validation` / `test` | 8,645 / 506 / 512 | same train fold; holdouts rated ≥ 7 |
 
-Data files live under `data/<config>/<split>-00000-of-00001.parquet` (plus CSV twins for convenience).
+Data files live under `data/<config>/<split>-00000-of-00001.parquet` (plus CSV twins). Paper-named copies of the canonical split are `benchmark/train.csv`, `benchmark/val.csv`, and `benchmark/test.csv`.
 
 ## Load with 🤗 Datasets
 
 ```python
 from datasets import load_dataset
 
-# default subset = interactions
-ds = load_dataset("MinhDS/ViHoRec")
-print(ds["train"][0])
-
 interactions = load_dataset("MinhDS/ViHoRec", "interactions")
 users = load_dataset("MinhDS/ViHoRec", "users")
 hotels = load_dataset("MinhDS/ViHoRec", "hotels")
-benchmark = load_dataset("MinhDS/ViHoRec", "benchmark")  # train + test
+content = load_dataset("MinhDS/ViHoRec", "content")
+benchmark = load_dataset("MinhDS/ViHoRec", "benchmark")  # train + validation + test
 
-# streaming (no full download)
-stream = load_dataset("MinhDS/ViHoRec", "interactions", split="train", streaming=True)
-for row in stream.take(3):
-    print(row)
+train = benchmark["train"]
+val = benchmark["validation"]
+test = benchmark["test"]
 ```
 
-Or with pandas / Polars:
+Or with pandas:
 
 ```python
 import pandas as pd
 
 interactions = pd.read_parquet("hf://datasets/MinhDS/ViHoRec/data/interactions/train-00000-of-00001.parquet")
 train = pd.read_parquet("hf://datasets/MinhDS/ViHoRec/data/benchmark/train-00000-of-00001.parquet")
+val = pd.read_parquet("hf://datasets/MinhDS/ViHoRec/data/benchmark/validation-00000-of-00001.parquet")
 test = pd.read_parquet("hf://datasets/MinhDS/ViHoRec/data/benchmark/test-00000-of-00001.parquet")
 ```
 
-## Repository layout (Hub)
+## Repository layout
 
 ```
 data/
 ├── interactions/train-00000-of-00001.{parquet,csv}
 ├── users/train-00000-of-00001.{parquet,csv}
 ├── hotels/train-00000-of-00001.{parquet,csv}
-└── benchmark/
-    ├── train-00000-of-00001.{parquet,csv}
-    └── test-00000-of-00001.{parquet,csv}
-benchmark/                   # reference artifacts (not in Viewer configs)
+├── content/train-00000-of-00001.{parquet,csv}
+├── benchmark/{train,validation,test}-00000-of-00001.{parquet,csv}
+└── benchmark_ge7/{train,validation,test}-00000-of-00001.{parquet,csv}
+benchmark/                   # reference artifacts (not all in Viewer configs)
+├── train.csv / val.csv / test.csv
 ├── user_map.csv / item_map.csv
 ├── split_config.json
-└── baseline_results*.csv
+├── PROTOCOL.md
+├── baseline_results*.csv / baseline_ci.md
+├── neural_results.csv
+├── recommenders_sota.md / cornac_sota.md
+└── relevance_ge_7/          # companion split, paper filenames
 DATASHEET.md
 LICENSE
 README.md
@@ -171,22 +240,38 @@ README.md
 
 ## Key statistics
 - Raw interactions: 18,274 (Booking 7,597 / Traveloka 6,273 / Ivivu 4,404)
-- After cleaning: **18,267** interactions, **6,832** users, **560** hotels
-- Entity matching merged 21 cross-site name variants; 78 hotels on ≥2 sites
-- Benchmark split: 800 users × 535 items, 9,787 train / 800 test, 97.53% sparse
+- After cleaning: **17,911** interactions, **6,822** reviewer keys, **560** hotels
+  (7 exact duplicates and 356 placeholder-name rows removed)
+- Released source rows: Booking 7,239 / Traveloka 6,273 / Ivivu 4,399
+- City- and type-aware entity resolution; 81 hotels on ≥2 sites; 0 hotels left with a conflicting location
+- Benchmark: 798 keys × 535 hotels, 8,645 train / 798 validation / 798 test, 97.60% sparse
+- Shortest train history = 2 (183 keys). Single-interaction keys stay in the corpus and are not scored
+- Closed-form test Recall@10 (validation-tuned): UserKNN 0.129, Hybrid-fixed (λ=0.7) 0.135, Adaptive Hybrid 0.133, EASE (λ=1000) 0.120
+- Sequential reference on the same freeze: SASRec and SSEPT Recall@10 = 0.172
+- Do not cite the submitted two-way Adaptive Hybrid Recall@10 of 0.1388
 
-## Full pipeline & code
-The reproducible construction scripts live on GitHub:
+## Full pipeline and code
+Construction scripts live on GitHub:
 [MinhNguyenDS/ViHoRec](https://github.com/MinhNguyenDS/ViHoRec).
 
 ```bash
-# rebuild Hub-ready parquet/csv layout from release/
-python scripts/prepare_hf_hub.py
-# then: huggingface-cli upload MinhDS/ViHoRec hf_hub/ . --repo-type dataset
+huggingface-cli upload MinhDS/ViHoRec hf_hub_r1/ . --repo-type dataset
 ```
 
 ## Citation
 
-If you use ViHoRec, please cite [arXiv:2607.12946](https://arxiv.org/abs/2607.12946)
+If you use ViHoRec, please cite [arXiv:2607.12946](https://arxiv.org/abs/2607.12946):
+
+```bibtex
+@misc{nguyen2026vihorec,
+  title={ViHoRec: A Quality-Controlled Vietnamese Hotel Recommendation Dataset and Short-History Benchmark},
+  author={Nguyen, Minh Hoang and Huynh, Tin Van and Nguyen, Kiet Van},
+  year={2026},
+  eprint={2607.12946},
+  archivePrefix={arXiv},
+  primaryClass={cs.IR},
+  url={https://arxiv.org/abs/2607.12946}
+}
+```
 
 See `DATASHEET.md` for provenance, ethics, and known limitations.

benchmark/PROTOCOL.md

@@ -0,0 +1,63 @@
+# ViHoRec public evaluation protocol
+
+This is the frozen protocol for the revised release. Scripts that search a
+hyper-parameter must read `val.csv`; `test.csv` is scored once.
+
+## Task name
+
+**Short-history / temporal leave-one-out recommendation.**
+
+It is not a strict user-cold-start protocol. Eligible users have at least four
+interactions in the cleaned corpus. After the last and second-last reviews are
+held out, every evaluated user still has at least two training interactions.
+The shortest test bucket is therefore train-history length **2**, not 0.
+
+The full corpus *is* cold-start dominated as a **dataset property**: most
+reviewer keys have a single interaction and are excluded from this split.
+That fact belongs in the characterization table, not in the name of the
+benchmark.
+
+## Fold assignment (per user, chronological)
+
+| Fold | Assignment | File |
+|---|---|---|
+| train | all interactions except the last two | `train.csv` |
+| val | second-last interaction | `val.csv` |
+| test | last interaction | `test.csv` |
+
+Ordering is per user, not a global calendar cut. One user's training review
+can be later than another user's test review. Identifiers are remapped to
+contiguous integers in `user_map.csv` / `item_map.csv`.
+
+Current counts (see `split_config.json`):
+
+- 798 users × 535 hotels
+- 8,645 train / 798 val / 798 test
+- train-history min / median / max = 2 / 5 / 208
+- 183 users have exactly two training interactions
+- sparsity 97.60%
+
+## Relevance (I10)
+
+The task is implicit next-item ranking. For a user, the single relevant item
+is the hotel they reviewed next, **regardless of the numeric score**. Every
+crawled rating is a positive (Cornac `rating_threshold = 0.5`). A 3/10 visit
+counts the same as a 10/10 visit: the signal is the visit.
+
+A sensitivity variant that keeps only holdouts rated ≥ 7 lives in
+`relevance_ge_7/` (train unchanged; 506 val / 512 test users remain). Do not
+overwrite the canonical files with that ablation.
+
+## What must not happen
+
+- Selecting λ, `(α_min, n0, τ)`, EASE λ, or neural hyper-parameters on `test.csv`
+- Calling the min-k = 4 protocol a cold-start benchmark
+- Quoting submitted Adaptive Recall@10 0.1388 / 0.141 or `adaptive_hybrid_results.md` from the two-way run (the val-tuned file now matches `baseline_ci.md`)
+
+## Statistics (Phase 3)
+
+Primary claims use per-user Recall@10 and NDCG@10. Uncertainty is a percentile
+bootstrap 95% CI over users (10,000 resamples). Model comparisons are paired
+Wilcoxon signed-rank tests on per-user Recall@10 (ties dropped). The pairing
+unit is the user: leave-one-out ranking scores are bounded and zero-inflated,
+so a Gaussian t-test is the wrong default.

benchmark/adaptive_hybrid_results.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d97548f9bd537d9076454ffaf5e698ef392dc573bb91f0c558edeb7b4a482bd
+size 580

benchmark/adaptive_hybrid_results.md

@@ -0,0 +1,8 @@
+| Method | MRR | MAP@5 | NDCG@5 | Precision@5 | Recall@5 | MAP@10 | NDCG@10 | Precision@10 | Recall@10 |
+|---|---|---|---|---|---|---|---|---|---|
+| UserKNN-cosine | 0.0606 | 0.0387 | 0.0492 | 0.0163 | 0.0815 | 0.0449 | 0.0644 | 0.0129 | 0.1291 |
+| Content-TFIDF | 0.0287 | 0.0126 | 0.0160 | 0.0053 | 0.0263 | 0.0164 | 0.0252 | 0.0055 | 0.0551 |
+| Hybrid-fixed (lam=0.7) | 0.0608 | 0.0378 | 0.0458 | 0.0140 | 0.0702 | 0.0462 | 0.0665 | 0.0135 | 0.1353 |
+| AdaptiveHybrid-uncon | 0.0512 | 0.0316 | 0.0375 | 0.0110 | 0.0551 | 0.0367 | 0.0499 | 0.0094 | 0.0940 |
+| AdaptiveHybrid-default | 0.0629 | 0.0404 | 0.0502 | 0.0160 | 0.0802 | 0.0471 | 0.0666 | 0.0132 | 0.1316 |
+| AdaptiveHybrid-valtuned | 0.0639 | 0.0405 | 0.0490 | 0.0150 | 0.0752 | 0.0480 | 0.0675 | 0.0133 | 0.1328 |

benchmark/baseline_ci.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef3cafef8d5b49f25af737d32a38a7039f87ea2d7f696bf9e9024a3493cfb8f5
+size 892

benchmark/baseline_ci.md

@@ -0,0 +1,17 @@
+# ViHoRec baselines (three-way split; val-tuned, test once)
+
+Hybrid-fixed λ = 0.7. Adaptive Hybrid {'alpha_min': 0.85, 'n0': 6.0, 'tau': 1.0} (val Recall@10 = 0.1341). EASE λ = 1000.
+
+95% CIs are percentile bootstrap over users (10,000 resamples). p is two-sided Wilcoxon signed-rank on per-user Recall@10 vs UserKNN. \* p<0.05, \*\* p<0.01, \*\*\* p<0.001.
+
+| Method | R@10 | 95% CI | N@10 | 95% CI | MRR | vs UserKNN |
+|---|---:|---|---:|---|---:|---|
+| Random | 0.0159 | [0.0109, 0.0213] | 0.0080 | [0.0053, 0.0109] | 0.0123 | 0.0000*** |
+| BPR-MF | 0.0977 | [0.0827, 0.1132] | 0.0485 | [0.0403, 0.0574] | 0.0493 | 0.0000*** |
+| MostPop | 0.1103 | [0.0890, 0.1328] | 0.0539 | [0.0425, 0.0663] | 0.0496 | 0.0588 |
+| ItemKNN-cosine | 0.0865 | [0.0677, 0.1065] | 0.0420 | [0.0321, 0.0527] | 0.0426 | 0.0001*** |
+| UserKNN-cosine | 0.1291 | [0.1065, 0.1529] | 0.0644 | [0.0519, 0.0777] | 0.0606 | — |
+| Content-TFIDF | 0.0551 | [0.0401, 0.0714] | 0.0252 | [0.0177, 0.0336] | 0.0287 | 0.0000*** |
+| Hybrid-fixed (lam=0.7) | 0.1353 | [0.1115, 0.1591] | 0.0665 | [0.0536, 0.0800] | 0.0608 | 0.5002 |
+| AdaptiveHybrid | 0.1328 | [0.1103, 0.1566] | 0.0675 | [0.0545, 0.0816] | 0.0639 | 0.4386 |
+| EASE (lam=1000) | 0.1203 | [0.0977, 0.1441] | 0.0595 | [0.0473, 0.0723] | 0.0573 | 0.2087 |

benchmark/baseline_results.csv

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:47088c309345f21459b3249a5b4106f221e7ec76e6a56cce5293bbefefda3b14
-size 529
+oid sha256:6297955bb8d525781dcbeefb58ca30e75b05fff2aef7ac7fcee26b463c9bbfa0
+size 767

benchmark/baseline_results.md

@@ -1,8 +1,11 @@
 | Method | MRR | MAP@5 | NDCG@5 | Precision@5 | Recall@5 | MAP@10 | NDCG@10 | Precision@10 | Recall@10 |
 |---|---|---|---|---|---|---|---|---|---|
-| Random | 0.0119 | 0.0036 | 0.0046 | 0.0016 | 0.0079 | 0.0054 | 0.0093 | 0.0023 | 0.0225 |
-| MostPop | 0.0496 | 0.0310 | 0.0385 | 0.0123 | 0.0612 | 0.0368 | 0.0528 | 0.0106 | 0.1062 |
-| ItemKNN-cosine | 0.0401 | 0.0212 | 0.0277 | 0.0095 | 0.0475 | 0.0252 | 0.0376 | 0.0079 | 0.0788 |
-| UserKNN-cosine | 0.0630 | 0.0387 | 0.0465 | 0.0140 | 0.0700 | 0.0472 | 0.0671 | 0.0134 | 0.1338 |
-| BPR-MF | 0.0512 | 0.0297 | 0.0369 | 0.0119 | 0.0592 | 0.0358 | 0.0519 | 0.0106 | 0.1058 |
-| Content-TFIDF | 0.0275 | 0.0118 | 0.0165 | 0.0063 | 0.0312 | 0.0153 | 0.0249 | 0.0057 | 0.0575 |
+| Random | 0.0123 | 0.0047 | 0.0059 | 0.0019 | 0.0096 | 0.0056 | 0.0080 | 0.0016 | 0.0159 |
+| BPR-MF | 0.0493 | 0.0284 | 0.0347 | 0.0109 | 0.0543 | 0.0340 | 0.0485 | 0.0097 | 0.0978 |
+| MostPop | 0.0496 | 0.0310 | 0.0390 | 0.0128 | 0.0639 | 0.0371 | 0.0539 | 0.0110 | 0.1103 |
+| ItemKNN-cosine | 0.0426 | 0.0245 | 0.0323 | 0.0113 | 0.0564 | 0.0285 | 0.0420 | 0.0086 | 0.0865 |
+| UserKNN-cosine | 0.0606 | 0.0387 | 0.0492 | 0.0163 | 0.0815 | 0.0449 | 0.0644 | 0.0129 | 0.1291 |
+| Content-TFIDF | 0.0287 | 0.0126 | 0.0160 | 0.0053 | 0.0263 | 0.0164 | 0.0252 | 0.0055 | 0.0551 |
+| Hybrid-fixed (lam=0.7) | 0.0608 | 0.0378 | 0.0458 | 0.0140 | 0.0702 | 0.0462 | 0.0665 | 0.0135 | 0.1353 |
+| AdaptiveHybrid | 0.0639 | 0.0405 | 0.0490 | 0.0150 | 0.0752 | 0.0480 | 0.0675 | 0.0133 | 0.1328 |
+| EASE (lam=1000) | 0.0573 | 0.0347 | 0.0434 | 0.0140 | 0.0702 | 0.0412 | 0.0595 | 0.0120 | 0.1203 |

benchmark/baseline_results_std.csv

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3395e73746e85b8f73b0ddbbbc0bdc11664c77e6408bbc057dc1fba8d4e4bd8a
-size 219
+oid sha256:1c1130d878d090f2cab4e75aea7723f8007f9df1fc14dc6f39cfb456793affcf
+size 216

benchmark/cornac_sota.md

@@ -0,0 +1,31 @@
+# Cornac generative / CF baselines (public three-way split)
+
+Fit on `train.csv`. Best-on-val checkpoint when the model supports it. `test.csv` scored once. Item text = hotel name + city + Facilities/Around/Vicinity/Price, encoded with `paraphrase-multilingual-MiniLM-L12-v2`.
+
+Sequential prefixes follow Cornac leave-last-out: the validation review is in the test-time history but is not a training label. That is one extra observed item relative to GRU4Rec in `run_neural_baselines.py` (train history only).
+
+SANSA is item–item CF (RecSys 2023). It uses the same train-only history and seen-item mask as EASE / UserKNN. SuiteSparse/CHOLMOD cannot build on this Windows env, and Cornac SANSA ICF is unstable on the 19.5%-dense 535-item gramian (the 150k-item sparse path). Reported SANSA is the exact dense LDL — the CHOLMOD density=1 closed form of Spišák et al. at this catalog size. Val-tuned λ=1000.
+
+| Method | MRR | NDCG@5 | Recall@5 | NDCG@10 | Recall@10 | Train (s) | Test (s) |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| TIGER | 0.0599 | 0.0495 | 0.0802 | 0.0734 | 0.1566 | 19974 | 155 |
+| RPG | 0.0673 | 0.0515 | 0.0802 | 0.0722 | 0.1454 | 2299 | 35 |
+| LETTER | 0.0569 | 0.0484 | 0.0802 | 0.0698 | 0.1466 | 28724 | 171 |
+| SANSA | 0.0583 | 0.0450 | 0.0727 | 0.0598 | 0.1190 | 2 | 0 |
+
+95% bootstrap CI over users; Wilcoxon on per-user Recall@10 vs UserKNN (R@10 = 0.1291). \* p<0.05.
+
+| Method | R@10 | 95% CI | N@10 | vs UserKNN |
+|---|---:|---|---:|---|
+| TIGER | 0.1566 | [0.1316, 0.1830] | 0.0734 | 0.0233* |
+| RPG | 0.1454 | [0.1216, 0.1704] | 0.0722 | 0.2596 |
+| LETTER | 0.1466 | [0.1228, 0.1717] | 0.0698 | 0.2050 |
+| SANSA | 0.1190 | [0.0977, 0.1416] | 0.0598 | 0.1441 |
+
+Reference (same freeze, train-only history): UserKNN 0.1291, Hybrid-fixed 0.1353, GRU4Rec 0.1604, EASE λ=1000 0.1203.
+
+## Skipped
+
+- **Companion**: Needs SentimentModality (aspect–opinion tuples from reviews). ViHoRec does not release review text.
+- **HypAR**: Review-hypergraph model; the public release has no review documents.
+- **DiffGRM**: Paper-scale GPU recipe (d_model=256, 100 epochs, beam 128). This environment is torch CPU-only.

benchmark/cornac_sota_ci.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65f533fff80073e383943df316f7274cb38471b378b86696625743d770a2bb69
+size 423

benchmark/cornac_sota_per_user.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad909d285e2fa13a4f9ae497ccb0038cfbaa1de2719b340127a55d6761b51cb0
+size 34722

benchmark/cornac_sota_results.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9298c343d83a56f58eb2369542c5a65694399a2e2f3498302be72894fb013871
+size 414

benchmark/hybrid_lambda_sweep.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf1b53ae79c7dd847ff20f87c122d9d3d0e7d3dfb336c5ac4dcc8ecfc52a2510
+size 1158

benchmark/hybrid_sensitivity.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e417da92f9bc2d9dfca7d6c94ca55119e3460262e238bfca69dcdf97e6ad983
+size 2234

benchmark/item_map.csv

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a0719dd721ee53fc831096b97902346bd9d4d13e3099c1b49460c1b1c8f86f99
+oid sha256:45dfca77094b4a09deee9148474cbb52334d8837ab1f79c7a852551d4ec8aae6
 size 5792

benchmark/neural_results.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f09abe19325341702a46d4e7f75e15782c7a5fb153e8b834ea55c860feeff699
+size 298

benchmark/pairwise_tests.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e1ce76589d288a2a54f5044b9f5fa10c1113eb917ed77afc01de10bcb81845b
+size 784

benchmark/per_user_metrics.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d17e3849418ab272c1cecad490fffd1862f6118d1f4f48e046c73714e140fa63
+size 74471

benchmark/recommenders_sota.md

@@ -0,0 +1,29 @@
+# Recommenders-team baselines (public three-way split)
+
+Library: [recommenders-team/recommenders](https://github.com/recommenders-team/recommenders). The repo does not contain 2025–2026 *papers*; NEWS last updated April 2025. The models below are the newest **implementations** on `main`: PyTorch SASRec / SSEPT (UniRec port, 2026) and PyTorch NeuMF.
+
+Protocol matches `run_neural_baselines.py`: train on `train.csv`, select epoch on `val.csv`, score `test.csv` once. Train-only history, full 535-item catalogue, mask seen train items. Recommenders' built-in sampled HR@10 (100 negatives) is **not** used.
+
+| Method | MRR | NDCG@5 | Recall@5 | NDCG@10 | Recall@10 | Train (s) | Test (s) |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| SASRec | 0.0753 | 0.0580 | 0.0890 | 0.0845 | 0.1717 | 75 | 0 |
+| NCF-NeuMF | 0.0681 | 0.0529 | 0.0840 | 0.0726 | 0.1441 | 53 | 2 |
+| SSEPT | 0.0706 | 0.0554 | 0.0915 | 0.0810 | 0.1717 | 85 | 1 |
+
+95% bootstrap CI over users; Wilcoxon on per-user Recall@10 vs UserKNN (R@10 = 0.1291). \* p<0.05.
+
+| Method | R@10 | 95% CI | N@10 | vs UserKNN |
+|---|---:|---|---:|---|
+| SASRec | 0.1717 | [0.1466, 0.1980] | 0.0845 | 0.0007*** |
+| NCF-NeuMF | 0.1441 | [0.1203, 0.1692] | 0.0726 | 0.1396 |
+| SSEPT | 0.1717 | [0.1454, 0.1980] | 0.0810 | 0.0003*** |
+
+Reference (same freeze, train-only history): UserKNN 0.1291, GRU4Rec 0.1604, LightGCN 0.1454, NeuMF 0.1228, TIGER 0.1566.
+
+## Skipped
+
+- **LightGCN**: Already evaluated in-repo (neural_results.csv, test R@10=0.1454). Recommenders LightGCN is the same 2020 model on the DeepRec yaml stack.
+- **SUM**: TensorFlow DeepRec sequential (Lian et al., 2021). Not on the 2026 PyTorch UniRec line; extra TF dependency on Python 3.13.
+- **SLi-Rec**: TensorFlow DeepRec (Microsoft, 2019). Same TF stack as SUM.
+- **NextItNet**: TensorFlow DeepRec dilated CNN (Yuan et al., 2019).
+- **NRMS/NAML/LSTUR**: News recommenders; they need article text. ViHoRec has hotel metadata, not news bodies.

benchmark/recommenders_sota_ci.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6322ff490d6e8630b5d22cb8be2e415ba53593fa1c5cd36440c7e0a66418cfb4
+size 367

benchmark/recommenders_sota_per_user.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:434cbfbb1b20f6e45339e8bdbcad2ad8a0e23e717b3fb91185151f8e72952fb0
+size 27516

benchmark/recommenders_sota_results.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e89628085cd3b98ecdc24b7b041d45b8c4f82eb1d885097aea301a1142499927
+size 340

benchmark/relevance_ge_7/item_map.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45dfca77094b4a09deee9148474cbb52334d8837ab1f79c7a852551d4ec8aae6
+size 5792

benchmark/relevance_ge_7/split_config.json

@@ -0,0 +1,63 @@
+{
+  "protocol": "temporal leave-one-out, three-way (train / val / test)",
+  "fold_assignment": {
+    "test": "last interaction per user",
+    "val": "second-last interaction per user",
+    "train": "all earlier interactions"
+  },
+  "temporal_scope": {
+    "ordering": "per user, not global",
+    "note": "Each user's own history is split chronologically, so no user's future leaks into their training data. Across users the folds overlap in wall-clock time: one user's training review can be later than another user's test review. This is the standard leave-last-one-out protocol; a single global cut-off date would be a different and stricter setting, and is not used here."
+  },
+  "relevance": {
+    "definition": "implicit next-item: the chronologically next hotel the user reviewed",
+    "all_ratings_are_positive": false,
+    "relevance_threshold": 7.0,
+    "note": "Ratings are not thresholded by default; the holdout item is relevant because it was visited, not because it scored well."
+  },
+  "min_interactions": 4,
+  "seed": 42,
+  "n_users": 798,
+  "n_items": 535,
+  "n_train": 8645,
+  "n_val": 506,
+  "n_test": 512,
+  "train_history_length": {
+    "min": 2,
+    "median": 5.0,
+    "max": 208
+  },
+  "users_with_train_history_2": 183,
+  "users_evaluated": {
+    "val": 506,
+    "test": 512
+  },
+  "sparsity_pct": 97.6012,
+  "output_dir": "benchmark\\relevance_ge_7",
+  "folds": {
+    "train": {
+      "n": 8645,
+      "n_users": 798,
+      "n_items": 508,
+      "date_min": "2012-01-08",
+      "date_max": "2023-12-06",
+      "rating_mean": 7.5283
+    },
+    "val": {
+      "n": 506,
+      "n_users": 506,
+      "n_items": 232,
+      "date_min": "2015-10-20",
+      "date_max": "2023-12-07",
+      "rating_mean": 8.8534
+    },
+    "test": {
+      "n": 512,
+      "n_users": 512,
+      "n_items": 226,
+      "date_min": "2017-07-08",
+      "date_max": "2023-12-09",
+      "rating_mean": 8.782
+    }
+  }
+}

benchmark/relevance_ge_7/test.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef6d39f2c6580ea60aff7db836833aa5e62edce95808cc2281bd4cee0bd614cc
+size 11745

benchmark/relevance_ge_7/train.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:781a3b99e5a020179e8f1c427c2acf858d1035385040114f88e76ab1320c915e
+size 196724

benchmark/relevance_ge_7/user_map.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:053f3b8b0568c8e450b2196ba1c2433273b9d1a8dbcccd970729342dc20a0b1b
+size 15068

benchmark/relevance_ge_7/val.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ae5d879f9a6ca7493afd701fb05666ab6bdb9dd9bb883d78aff6e8a8ee2d62a
+size 11623

benchmark/split_config.json

@@ -1,10 +1,63 @@
 {
-  "protocol": "leave-last-one-out (temporal)",
+  "protocol": "temporal leave-one-out, three-way (train / val / test)",
+  "fold_assignment": {
+    "test": "last interaction per user",
+    "val": "second-last interaction per user",
+    "train": "all earlier interactions"
+  },
+  "temporal_scope": {
+    "ordering": "per user, not global",
+    "note": "Each user's own history is split chronologically, so no user's future leaks into their training data. Across users the folds overlap in wall-clock time: one user's training review can be later than another user's test review. This is the standard leave-last-one-out protocol; a single global cut-off date would be a different and stricter setting, and is not used here."
+  },
+  "relevance": {
+    "definition": "implicit next-item: the chronologically next hotel the user reviewed",
+    "all_ratings_are_positive": true,
+    "relevance_threshold": null,
+    "note": "Ratings are not thresholded by default; the holdout item is relevant because it was visited, not because it scored well."
+  },
   "min_interactions": 4,
   "seed": 42,
-  "n_users": 800,
+  "n_users": 798,
   "n_items": 535,
-  "n_train": 9787,
-  "n_test": 800,
-  "sparsity_pct": 97.5264
+  "n_train": 8645,
+  "n_val": 798,
+  "n_test": 798,
+  "train_history_length": {
+    "min": 2,
+    "median": 5.0,
+    "max": 208
+  },
+  "users_with_train_history_2": 183,
+  "users_evaluated": {
+    "val": 798,
+    "test": 798
+  },
+  "sparsity_pct": 97.6012,
+  "output_dir": "benchmark",
+  "folds": {
+    "train": {
+      "n": 8645,
+      "n_users": 798,
+      "n_items": 508,
+      "date_min": "2012-01-08",
+      "date_max": "2023-12-06",
+      "rating_mean": 7.5283
+    },
+    "val": {
+      "n": 798,
+      "n_users": 798,
+      "n_items": 289,
+      "date_min": "2015-10-20",
+      "date_max": "2023-12-07",
+      "rating_mean": 7.5561
+    },
+    "test": {
+      "n": 798,
+      "n_users": 798,
+      "n_items": 278,
+      "date_min": "2017-07-08",
+      "date_max": "2023-12-09",
+      "rating_mean": 7.4976
+    }
+  }
 }

benchmark/test.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb2d913269c597d986e242b65c53f7067f765c68915005054fae99426e97afbc
+size 18217

benchmark/train.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:781a3b99e5a020179e8f1c427c2acf858d1035385040114f88e76ab1320c915e
+size 196724

benchmark/user_map.csv

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c830948d8aa0afc75ea53f2a3005ae1ae3176ea57e653ddccf58271c68fd3716
-size 15106
+oid sha256:053f3b8b0568c8e450b2196ba1c2433273b9d1a8dbcccd970729342dc20a0b1b
+size 15068

benchmark/val.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dcaeba814337360e1c483be1e3d61201ecbfb587555938114ae907bf8a15b948
+size 18229

data/benchmark/test-00000-of-00001.csv

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ab7ea148c81fb5305521ba317ffbb66ab598b37566e68161240fa2fe684a51b2
-size 16666
+oid sha256:eb2d913269c597d986e242b65c53f7067f765c68915005054fae99426e97afbc
+size 18217

data/benchmark/test-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:96a7b7146bc32b1a0dc153646e12360b32c866613d431a317523bbf9adf13ffc
-size 13009
+oid sha256:da63b684ed22429f54b5337bf4dc5ba0ad394ee1c8cef5981c87ee05a7c6ff8c
+size 13372

data/benchmark/train-00000-of-00001.csv

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ff236a8d562028ff9d2c21cbc159225d84a1e6d06c7983e00310c9eca7fade53
-size 203527
+oid sha256:781a3b99e5a020179e8f1c427c2acf858d1035385040114f88e76ab1320c915e
+size 196724

data/benchmark/train-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:90b391f355d11ef7c56a563a4260a0cee723999f24883079a06bc08d408e18c5
-size 59180
+oid sha256:22da58f3157f2f4d6776b5a0ddb1763f3f3a31d0230bb1fa0fcb64fcc49aec1f
+size 55033

data/benchmark/validation-00000-of-00001.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dcaeba814337360e1c483be1e3d61201ecbfb587555938114ae907bf8a15b948
+size 18229

data/benchmark/validation-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e179b599b3124ea6eb41e811e69eaddf1586c415d3eaaede9c2f4af5f60c870
+size 14263

data/benchmark_ge7/test-00000-of-00001.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef6d39f2c6580ea60aff7db836833aa5e62edce95808cc2281bd4cee0bd614cc
+size 11745

data/benchmark_ge7/test-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f842d26b4037559a145ed497fbcff1f8ca6dfab0e117f09984702452e51b5ea
+size 9768

data/benchmark_ge7/train-00000-of-00001.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:781a3b99e5a020179e8f1c427c2acf858d1035385040114f88e76ab1320c915e
+size 196724

data/benchmark_ge7/train-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22da58f3157f2f4d6776b5a0ddb1763f3f3a31d0230bb1fa0fcb64fcc49aec1f
+size 55033

data/benchmark_ge7/validation-00000-of-00001.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ae5d879f9a6ca7493afd701fb05666ab6bdb9dd9bb883d78aff6e8a8ee2d62a
+size 11623

data/benchmark_ge7/validation-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44330721be7b56fdb2d9055fc736c402854994b367b7350d927ff3ea928b904d
+size 10280

data/content/train-00000-of-00001.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b0f7871f42f561590a04d2414d2fecdde0610df6a00a66c223b0639aead9927
+size 120172

data/content/train-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40c1d8b18cf09d44804d46c183d935e9bf60352ccc221e68946f008d0fee5186
+size 35735