Training products derived from islamlab/islamic-sciences
Browse files- README.md +11 -11
- attribution/train-00000.parquet +2 -2
- cpt/train-00000.parquet +2 -2
- eval_attribution/test-00000.parquet +2 -2
- eval_retrieval/test-00000.parquet +2 -2
- retrieval/train-00000.parquet +2 -2
README.md
CHANGED
|
@@ -58,16 +58,16 @@ model.
|
|
| 58 |
|
| 59 |
| config | split | rows | what it is |
|
| 60 |
|---|---|---:|---|
|
| 61 |
-
| `cpt` | train | 380,
|
| 62 |
-
| `retrieval` | train |
|
| 63 |
-
| `attribution` | train |
|
| 64 |
| `eval_retrieval` | test | 14,369 | Same shape, held-out works only |
|
| 65 |
-
| `eval_attribution` | test |
|
| 66 |
|
| 67 |
## The held-out split
|
| 68 |
|
| 69 |
The holdout is at the level of the **work**, not the row: 76 of the
|
| 70 |
-
2,
|
| 71 |
and no line of their text appears in `cpt`, `retrieval` or `attribution`. A
|
| 72 |
model trained on the train configs has therefore never read the books it is
|
| 73 |
evaluated on. The assignment is deterministic, so it is the same for everyone.
|
|
@@ -76,11 +76,11 @@ evaluated on. The assignment is deterministic, so it is the same for everyone.
|
|
| 76 |
|
| 77 |
| task | question | rows (train) |
|
| 78 |
|---|---|---:|
|
| 79 |
-
| `author` | من مؤلف هذا النص؟ |
|
| 80 |
-
| `title` | من أي كتاب أُخذ هذا النص؟ |
|
| 81 |
-
| `science` | في أي علم يُصنَّف هذا الكتاب؟ |
|
| 82 |
-
| `century` | في أي قرن هجري توفي المؤلف؟ |
|
| 83 |
-
| `locus` | في أي جزء وصفحة يقع هذا النص؟ |
|
| 84 |
|
| 85 |
`messages` is a JSON-encoded chat turn — parse it with `json.loads`. `answer`
|
| 86 |
holds the exact expected string. The `locus` task deliberately includes the
|
|
@@ -107,7 +107,7 @@ negatives = json.loads(row["negatives"])
|
|
| 107 |
|
| 108 |
| Dataset | What it holds |
|
| 109 |
|---|---|
|
| 110 |
-
| [`islamlab/islamic-sciences`](https://huggingface.co/datasets/islamlab/islamic-sciences) | The corpus itself — eight sciences, 2,
|
| 111 |
| **`islamlab/islamic-sciences-training`** | The training products above |
|
| 112 |
| [`quranlab/*`](https://huggingface.co/quranlab) · [`risaleinur/*`](https://huggingface.co/risaleinur) | Qur'an and hadith; the Risale-i Nur corpus |
|
| 113 |
|
|
|
|
| 58 |
|
| 59 |
| config | split | rows | what it is |
|
| 60 |
|---|---|---:|---|
|
| 61 |
+
| `cpt` | train | 380,357 | Apparatus-free, deduplicated text with a provenance header — 1.55B characters |
|
| 62 |
+
| `retrieval` | train | 413,737 | A section heading as the query, a passage from that section as the positive, and hard negatives drawn from elsewhere in the same book |
|
| 63 |
+
| `attribution` | train | 65,415 | Citation questions with exact answers |
|
| 64 |
| `eval_retrieval` | test | 14,369 | Same shape, held-out works only |
|
| 65 |
+
| `eval_attribution` | test | 17,035 | Same shape, held-out works only |
|
| 66 |
|
| 67 |
## The held-out split
|
| 68 |
|
| 69 |
The holdout is at the level of the **work**, not the row: 76 of the
|
| 70 |
+
2,600 works are assigned to evaluation by a hash of their identifier,
|
| 71 |
and no line of their text appears in `cpt`, `retrieval` or `attribution`. A
|
| 72 |
model trained on the train configs has therefore never read the books it is
|
| 73 |
evaluated on. The assignment is deterministic, so it is the same for everyone.
|
|
|
|
| 76 |
|
| 77 |
| task | question | rows (train) |
|
| 78 |
|---|---|---:|
|
| 79 |
+
| `author` | من مؤلف هذا النص؟ | 13,083 |
|
| 80 |
+
| `title` | من أي كتاب أُخذ هذا النص؟ | 13,083 |
|
| 81 |
+
| `science` | في أي علم يُصنَّف هذا الكتاب؟ | 13,083 |
|
| 82 |
+
| `century` | في أي قرن هجري توفي المؤلف؟ | 13,083 |
|
| 83 |
+
| `locus` | في أي جزء وصفحة يقع هذا النص؟ | 13,083 |
|
| 84 |
|
| 85 |
`messages` is a JSON-encoded chat turn — parse it with `json.loads`. `answer`
|
| 86 |
holds the exact expected string. The `locus` task deliberately includes the
|
|
|
|
| 107 |
|
| 108 |
| Dataset | What it holds |
|
| 109 |
|---|---|
|
| 110 |
+
| [`islamlab/islamic-sciences`](https://huggingface.co/datasets/islamlab/islamic-sciences) | The corpus itself — eight sciences, 2,600 works |
|
| 111 |
| **`islamlab/islamic-sciences-training`** | The training products above |
|
| 112 |
| [`quranlab/*`](https://huggingface.co/quranlab) · [`risaleinur/*`](https://huggingface.co/risaleinur) | Qur'an and hadith; the Risale-i Nur corpus |
|
| 113 |
|
attribution/train-00000.parquet
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:dcdd43534fc2d157b7a328021d79e79c903eef8219643d0fb5faa1f9426b618e
|
| 3 |
+
size 6578217
|
cpt/train-00000.parquet
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b36e85cb38fe52c0b89cede5b5aa71b6f6c10edb800277c0b236763a1dd48cf6
|
| 3 |
+
size 639753664
|
eval_attribution/test-00000.parquet
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:badee030ea5a2c45cc4357c6b0c55a1ab8c45b513c0e718002cf220f00b58494
|
| 3 |
+
size 1762147
|
eval_retrieval/test-00000.parquet
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:879ed6da7175f7bd555b357be7e4f2a2f6fc71a7876f8ffde02c01cd59fbdb89
|
| 3 |
+
size 24243696
|
retrieval/train-00000.parquet
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4978dbba5ac6fb1e19d2e65fb3b39e9bc313c7846aea838f4099029bbccd7f32
|
| 3 |
+
size 577862756
|