Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -195,7 +195,8 @@ parallel_data_unfiltered/
|
|
| 195 |
...
|
| 196 |
|
| 197 |
With ALPHA being `0.005` or `None` and LENGTH_PENALTY_TYPE can be `"relative"`, `"absolute"` or `None`.
|
| 198 |
-
|
|
|
|
| 199 |
|
| 200 |
### File Organization
|
| 201 |
|
|
@@ -210,3 +211,7 @@ Same structure for `/parallel_data_filtered`.
|
|
| 210 |
| `embedded/filtered_notail_dedup/filtered_topleveldomain` | Embedding only with chosen TLDs|
|
| 211 |
| `fineweb_original/` | Original data
|
| 212 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 195 |
...
|
| 196 |
|
| 197 |
With ALPHA being `0.005` or `None` and LENGTH_PENALTY_TYPE can be `"relative"`, `"absolute"` or `None`.
|
| 198 |
+
|
| 199 |
+
For `/parallel_data_filtered` we chose to always include a penalty.
|
| 200 |
|
| 201 |
### File Organization
|
| 202 |
|
|
|
|
| 211 |
| `embedded/filtered_notail_dedup/filtered_topleveldomain` | Embedding only with chosen TLDs|
|
| 212 |
| `fineweb_original/` | Original data
|
| 213 |
|
| 214 |
+
|
| 215 |
+
## Future Work
|
| 216 |
+
Further improvements could explore the use of alternative ALPHA values and experiment with different filtering methods to refine data quality. Other directions include testing different penalty calculation types, incorporating other alignment approaches and applying domain-specific preprocessing to optimize results for different datasets and use cases.
|
| 217 |
+
|