Add val split (2000 queries carved from test, seed=42); drop source_split column

README.md

@@ -19,6 +19,8 @@ configs:
     data_files:
       - split: train
         path: default/train.parquet
+      - split: val
+        path: default/val.parquet
       - split: test
         path: default/test.parquet
 ---
@@ -33,20 +35,24 @@ from [`devrim/goodwiki_long`](https://huggingface.co/datasets/devrim/goodwiki_lo
 
 ## Layout
 
-| config    | split(s)        | rows   |
-|-----------|-----------------|--------|
-| `corpus`  | `corpus`        | 53,562 |
-| `queries` | `queries`       | 17,854 |
-| `default` | `train`, `test` | 30,000 / 23,562 |
+| config    | split(s)             | rows                            |
+|-----------|----------------------|---------------------------------|
+| `corpus`  | `corpus`             | 53,562                          |
+| `queries` | `queries`            | 17,854                          |
+| `default` | `train`, `val`, `test` | 30,000 / 6,000 / 17,562       |
 
-The corpus and queries are single shared pools. Train/test only exists in `default`
-(qrels), so a query in the train split competes against the *full* corpus and
-likewise for test — train synthetics serve as hard negatives for test queries and
-vice-versa.
+The corpus and queries are single shared pools. Splits exist only on `default`
+(qrels), so every query — train, val, or test — competes against the *full*
+corpus, and synthetics from other splits serve as hard negatives.
 
-The `source_split` column on `corpus` and `queries` records which split each row
-was generated under in the source datasets, in case downstream code wants to
-reconstruct the original disjoint pools.
+| qrels split | unique queries | qrels rows |
+|---|---|---|
+| `train` | 10,000 | 30,000 |
+| `val`   |  2,000 |  6,000 |
+| `test`  |  5,854 | 17,562 |
+
+Val was carved out of the original test split (random sample, seed=42) before
+upload. A query's split is determined by which qrels split it appears in.
 
 ## Configs
 
@@ -60,7 +66,6 @@ reconstruct the original disjoint pools.
 | `other_article_id` | string | id of the secondary article mixed in (only meaningful for `distractor`) |
 | `primary_topic` | string | source article title + short description |
 | `integration_topic` | string | secondary article title + short description |
-| `source_split` | string | `train` or `test` (from source dataset) |
 
 ### `queries`
 | field | type | notes |
@@ -70,7 +75,6 @@ reconstruct the original disjoint pools.
 | `title` | string | |
 | `description` | string | |
 | `categories` | list[string] | |
-| `source_split` | string | `train` or `test` (from source dataset) |
 
 ### `default` (qrels)
 | field | type | notes |
@@ -88,5 +92,6 @@ from datasets import load_dataset
 
 corpus  = load_dataset("devrim/goodwiki_long_synthetic_ir", "corpus",  split="corpus")
 queries = load_dataset("devrim/goodwiki_long_synthetic_ir", "queries", split="queries")
-qrels   = load_dataset("devrim/goodwiki_long_synthetic_ir", "default", split="test")
+val     = load_dataset("devrim/goodwiki_long_synthetic_ir", "default", split="val")
+test    = load_dataset("devrim/goodwiki_long_synthetic_ir", "default", split="test")
 ```

corpus/corpus.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c619e3cc62041b15753618dbb5d83df888e2cce963724b22ba4713c12e27958c
-size 285981831
+oid sha256:633c285a2a1835822436ae754373b59abed2666d9b4d4b848ed212cf6bd7b227
+size 285968412

default/test.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1a194e35211bc84a98ae4b577d01d7c3f671e6677351f41b7e0cfeabec1e3789
-size 329083
+oid sha256:747286d5fc8ee880bf2c18f2a782f0895e6fc9ee699cd4701b0b8ec5f7905815
+size 246876

default/val.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53e095dd3ffdc92e3821ea616771840cf39ab51de74a00fbfbd11ecb91fba713
+size 85663

queries/queries.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:78e8155f4d8ac4cd348483e11925a0fe689ab23b401612c93d2641207a447218
-size 327766343
+oid sha256:2d6ec04f53dd354d888b7f8ac5a5823ae0e58fdf348336180dbe2e33d754b0f7
+size 327761722