Kenneth Enevoldsen commited on
Commit
5cc0948
·
unverified ·
1 Parent(s): 4c55782

updated paper

Browse files
Files changed (2) hide show
  1. paper/figure_baseline.png +3 -0
  2. paper/paper.md +34 -0
paper/figure_baseline.png ADDED

Git LFS Details

  • SHA256: 6cf32e792c92d9188a0e5a02ca2ffabe2516e97eb3914f4c5641fe5a47d8a045
  • Pointer size: 131 Bytes
  • Size of remote file: 113 kB
paper/paper.md CHANGED
@@ -193,3 +193,37 @@ enviromental:
193
  - common codebase lead to less duplication of dataset and reduces storage required
194
  - continual ci running on large datasets could be a concern. Currently out tests use a total of XXX Co2-eq (estimated using codecarbon). however we have already seen people using training [@fineweb] and evaluating LLMs to appriximate dataset quality, such workflows could quickly incrase the co2 consumption.
195
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
193
  - common codebase lead to less duplication of dataset and reduces storage required
194
  - continual ci running on large datasets could be a concern. Currently out tests use a total of XXX Co2-eq (estimated using codecarbon). however we have already seen people using training [@fineweb] and evaluating LLMs to appriximate dataset quality, such workflows could quickly incrase the co2 consumption.
195
 
196
+
197
+
198
+
199
+
200
+ ## Aditional content
201
+
202
+ Comparison table
203
+
204
+
205
+ | | Size | Sufficient Documentation | Data availability | Legal Status | Quality |
206
+ | ---------------------- | ---- | ------------------------ | ----------------- | --------------- | -------------- |
207
+ | Danish Dynaword (Ours) | 3.5B | Replicable^ | Open Access | Openly Licensed | Mixed (high) |
208
+ | Danish Gigaword* | | Documentary | Open Access | Openly Licensed | Mixed (high) |
209
+ | Common Corpus (dan) | | Replicable | Open Access | Openly Licensed | OCR (low) |
210
+ | Fineweb (dan) | | Replicable | Open Access | | Mixed (medium) |
211
+
212
+
213
+
214
+ <!--
215
+ Could we create an interesting figure of this Marton? See figure 1
216
+ better notion of quality? Potentially a bit more objective?
217
+
218
+ -->
219
+
220
+ *The Danish gigaword subsection included in Danish Dynaword. I.e. the subsection that is permissibly licensed.
221
+ ^Some datasets are derived from Danish Gigaword, some of these subsection are not (currently) replicable
222
+
223
+ This follows the scheme from figure 1 (https://arxiv.org/abs/2501.08365)
224
+
225
+ Add comparison number of tokens comparison:
226
+ Common Corpus (DA) -
227
+ Gigaword (DA) - Open
228
+ M-Fineweb (DA) -
229
+ -->