Datasets:
Document parquet load for windowed_w8192_s4579 (json builder cannot mix formats)
Browse files
README.md
CHANGED
|
@@ -149,15 +149,21 @@ This dataset is **derived** from a specific tokenizer, window length, and step.
|
|
| 149 |
**Regenerate it** if any of those change. Never mix windows from different tokenizers
|
| 150 |
or strides in one training run.
|
| 151 |
|
| 152 |
-
Load the
|
|
|
|
|
|
|
|
|
|
| 153 |
|
| 154 |
```python
|
| 155 |
from datasets import load_dataset
|
| 156 |
|
| 157 |
ds = load_dataset(
|
| 158 |
-
"
|
| 159 |
-
|
| 160 |
-
|
|
|
|
|
|
|
|
|
|
| 161 |
)
|
| 162 |
```
|
| 163 |
|
|
|
|
| 149 |
**Regenerate it** if any of those change. Never mix windows from different tokenizers
|
| 150 |
or strides in one training run.
|
| 151 |
|
| 152 |
+
Load the windowed parquet shards — not the unpinned default book-level jsonl.
|
| 153 |
+
``load_dataset(repo, "windowed_w8192_s4579")`` uses the JSON builder because the
|
| 154 |
+
default README config is ``*.jsonl``; ``datasets`` infers one builder for the
|
| 155 |
+
whole repo from that first config.
|
| 156 |
|
| 157 |
```python
|
| 158 |
from datasets import load_dataset
|
| 159 |
|
| 160 |
ds = load_dataset(
|
| 161 |
+
"parquet",
|
| 162 |
+
data_files={
|
| 163 |
+
"train": "hf://datasets/karma689/tibetan-quotation-detection@w1.0/windowed_w8192_s4579/train/*.parquet",
|
| 164 |
+
"validation": "hf://datasets/karma689/tibetan-quotation-detection@w1.0/windowed_w8192_s4579/validation/*.parquet",
|
| 165 |
+
"test": "hf://datasets/karma689/tibetan-quotation-detection@w1.0/windowed_w8192_s4579/test/*.parquet",
|
| 166 |
+
},
|
| 167 |
)
|
| 168 |
```
|
| 169 |
|