| --- |
| license: apache-2.0 |
| --- |
| |
| # GRSB |
|
|
| GRSB (Gorse Reranker System Benchmark) provides recommendation datasets for evaluating rerankers with user histories, item documents, and binary feedback labels. |
|
|
| ## Data format |
|
|
| Each dataset is stored under `data/<dataset>/` with three files: |
|
|
| | File | Description | |
| | --- | --- | |
| | `users.jsonl` | One JSON object per user. | |
| | `items.jsonl` | One JSON object per item, including item metadata and text used to render documents. | |
| | `feedback.csv` | User-item interactions in `user_id,item_id,label,timestamp` format. `label` is `1` for positive feedback and `0` for negative feedback. | |
|
|
| Prompt templates are stored under `prompt/<dataset>/`: |
|
|
| | File | Description | |
| | --- | --- | |
| | `instruct.txt` | Reranker instruction. | |
| | `query.jinja2` | Template for rendering a user's history into a query. | |
| | `document.jinja2` | Template for rendering item metadata into a document. | |
|
|
| ## Dataset sizes |
|
|
| | Dataset | Domain | Users | Items | Feedback | Positive | Negative | Size | |
| | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | |
| | `ml-100k` | Movies | 943 | 1,682 | 100,000 | 55,375 | 44,625 | 2.70 MiB | |
| | `ml-1m` | Movies | 6,040 | 3,883 | 1,000,209 | 575,281 | 424,928 | 21.97 MiB | |
| | `spoilers` | Books | 18,892 | 25,475 | 1,378,033 | 883,573 | 494,460 | 156.61 MiB | |
|
|
| Notes: |
|
|
| - `ml-1m` has 3,883 movie metadata records, while its feedback file references 3,706 of those items. |
| - Movie item descriptions are generated from titles and stored in `items.jsonl` for document rendering. |
| - `spoilers` item metadata is larger because it includes richer book text and metadata. |
|
|
| ## File-level sizes |
|
|
| | Dataset | File | Rows | Size | |
| | --- | --- | ---: | ---: | |
| | `ml-100k` | `users.jsonl` | 943 | 0.07 MiB | |
| | `ml-100k` | `items.jsonl` | 1,682 | 0.74 MiB | |
| | `ml-100k` | `feedback.csv` | 100,000 | 1.89 MiB | |
| | `ml-1m` | `users.jsonl` | 6,040 | 0.42 MiB | |
| | `ml-1m` | `items.jsonl` | 3,883 | 0.96 MiB | |
| | `ml-1m` | `feedback.csv` | 1,000,209 | 20.59 MiB | |
| | `spoilers` | `users.jsonl` | 18,892 | 0.85 MiB | |
| | `spoilers` | `items.jsonl` | 25,475 | 84.42 MiB | |
| | `spoilers` | `feedback.csv` | 1,378,033 | 71.34 MiB | |
|
|
