Sudehsna commited on
Commit
0992d7e
·
verified ·
1 Parent(s): 9630031

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +6 -1
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
- Same structure for `/parallel_data_filtered`.
 
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
+