BramVanroy
commited on
Update README.md
Browse files
README.md
CHANGED
@@ -293,13 +293,13 @@ This is a combined and filtered version of [CulturaX](https://huggingface.co/dat
|
|
293 |
|
294 |
Different configs are available based on the number of tokens (see a section below with an overview). This can be useful if you want to know exactly how many tokens you have. Great for using as a streaming dataset, too. Tokens are counted as white-space tokens, so depending on your tokenizer, you'll likely end up with more tokens than indicated here.
|
295 |
|
296 |
-
Every config also has a test set (for validation) of 1% the total size of the dataset, minimally 1 max. 64k samples (~
|
297 |
|
298 |
Wikipedia and CulturaX were suffled before merging and the teset set creation was also shuffled. Priority is given to Wikipedia to prioritize knowledge-content, so the smaller configs will consist exclusively of Wikipedia and for the larger configs we augment with CulturaX. Every config builds further on the previous, so this means that every config contains the same data as the smaller ones and more HOWEVER their train/test splits are not the same, so test set of one config may overlap with samples for another training set. This is usually not a problem but just be aware that you do not train on one config's training set and test with another config's test set.
|
299 |
|
300 |
## Configs
|
301 |
|
302 |
-
### 10k -- 79 samples -- 10,087 tokens
|
303 |
- ratio_wikipedia: 100.00%
|
304 |
- total_num_tokens: 10,087
|
305 |
- train_num_tokens: 9,205
|
@@ -308,7 +308,7 @@ Wikipedia and CulturaX were suffled before merging and the teset set creation wa
|
|
308 |
- train_num_samples: 78
|
309 |
- test_num_samples: 1
|
310 |
|
311 |
-
### 100k -- 1,057 samples -- 100,075 tokens
|
312 |
- ratio_wikipedia: 100.00%
|
313 |
- total_num_tokens: 100,075
|
314 |
- train_num_tokens: 98,044
|
@@ -317,7 +317,7 @@ Wikipedia and CulturaX were suffled before merging and the teset set creation wa
|
|
317 |
- train_num_samples: 1,047
|
318 |
- test_num_samples: 10
|
319 |
|
320 |
-
### 1M -- 10,802 samples -- 1,000,239 tokens
|
321 |
- ratio_wikipedia: 100.00%
|
322 |
- total_num_tokens: 1,000,239
|
323 |
- train_num_tokens: 991,119
|
@@ -326,7 +326,7 @@ Wikipedia and CulturaX were suffled before merging and the teset set creation wa
|
|
326 |
- train_num_samples: 10,694
|
327 |
- test_num_samples: 108
|
328 |
|
329 |
-
### 10M -- 141,263 samples -- 10,000,022 tokens
|
330 |
- ratio_wikipedia: 100.00%
|
331 |
- total_num_tokens: 10,000,022
|
332 |
- train_num_tokens: 9,874,772
|
@@ -335,7 +335,7 @@ Wikipedia and CulturaX were suffled before merging and the teset set creation wa
|
|
335 |
- train_num_samples: 139,851
|
336 |
- test_num_samples: 1,412
|
337 |
|
338 |
-
### 100M -- 1,028,484 samples -- 100,000,047 tokens
|
339 |
- ratio_wikipedia: 100.00%
|
340 |
- total_num_tokens: 100,000,047
|
341 |
- train_num_tokens: 99,013,372
|
@@ -344,7 +344,7 @@ Wikipedia and CulturaX were suffled before merging and the teset set creation wa
|
|
344 |
- train_num_samples: 1,018,200
|
345 |
- test_num_samples: 10,284
|
346 |
|
347 |
-
### 1B -- 5,153,898 samples -- 1,000,000,187 tokens
|
348 |
- ratio_wikipedia: 61.21%
|
349 |
- total_num_tokens: 1,000,000,187
|
350 |
- train_num_tokens: 989,990,190
|
@@ -353,7 +353,7 @@ Wikipedia and CulturaX were suffled before merging and the teset set creation wa
|
|
353 |
- train_num_samples: 5,102,360
|
354 |
- test_num_samples: 51,538
|
355 |
|
356 |
-
### 5B -- 20,833,009 samples -- 5,000,000,076 tokens
|
357 |
- ratio_wikipedia: 25.35%
|
358 |
- total_num_tokens: 5,000,000,076
|
359 |
- train_num_tokens: 4,984,493,654
|
@@ -362,7 +362,7 @@ Wikipedia and CulturaX were suffled before merging and the teset set creation wa
|
|
362 |
- train_num_samples: 20,769,009
|
363 |
- test_num_samples: 64,000
|
364 |
|
365 |
-
### 10B -- 40,240,566 samples -- 10,000,000,115 tokens
|
366 |
- ratio_wikipedia: 18.41%
|
367 |
- total_num_tokens: 10,000,000,115
|
368 |
- train_num_tokens: 9,984,156,828
|
@@ -371,7 +371,7 @@ Wikipedia and CulturaX were suffled before merging and the teset set creation wa
|
|
371 |
- train_num_samples: 40,176,566
|
372 |
- test_num_samples: 64,000
|
373 |
|
374 |
-
### 15B -- 59,648,123 samples -- 15,000,000,154 tokens
|
375 |
- ratio_wikipedia: 15.98%
|
376 |
- total_num_tokens: 15,000,000,154
|
377 |
- train_num_tokens: 14,983,970,518
|
@@ -380,7 +380,7 @@ Wikipedia and CulturaX were suffled before merging and the teset set creation wa
|
|
380 |
- train_num_samples: 59,584,123
|
381 |
- test_num_samples: 64,000
|
382 |
|
383 |
-
### 20B -- 79,055,679 samples -- 20,000,000,009 tokens
|
384 |
- ratio_wikipedia: 14.75%
|
385 |
- total_num_tokens: 20,000,000,009
|
386 |
- train_num_tokens: 19,983,799,357
|
@@ -389,7 +389,7 @@ Wikipedia and CulturaX were suffled before merging and the teset set creation wa
|
|
389 |
- train_num_samples: 78,991,679
|
390 |
- test_num_samples: 64,000
|
391 |
|
392 |
-
### 25B -- 98,463,236 samples -- 25,000,000,048 tokens
|
393 |
- ratio_wikipedia: 14.00%
|
394 |
- total_num_tokens: 25,000,000,048
|
395 |
- train_num_tokens: 24,983,765,326
|
@@ -398,7 +398,7 @@ Wikipedia and CulturaX were suffled before merging and the teset set creation wa
|
|
398 |
- train_num_samples: 98,399,236
|
399 |
- test_num_samples: 64,000
|
400 |
|
401 |
-
### 30B -- 117,870,793 samples -- 30,000,000,087 tokens
|
402 |
- ratio_wikipedia: 13.50%
|
403 |
- total_num_tokens: 30,000,000,087
|
404 |
- train_num_tokens: 29,983,707,932
|
|
|
293 |
|
294 |
Different configs are available based on the number of tokens (see a section below with an overview). This can be useful if you want to know exactly how many tokens you have. Great for using as a streaming dataset, too. Tokens are counted as white-space tokens, so depending on your tokenizer, you'll likely end up with more tokens than indicated here.
|
295 |
|
296 |
+
Every config also has a test set (for validation) of 1% the total size of the dataset, minimally 1 max. 64k samples (~16M tokens).
|
297 |
|
298 |
Wikipedia and CulturaX were suffled before merging and the teset set creation was also shuffled. Priority is given to Wikipedia to prioritize knowledge-content, so the smaller configs will consist exclusively of Wikipedia and for the larger configs we augment with CulturaX. Every config builds further on the previous, so this means that every config contains the same data as the smaller ones and more HOWEVER their train/test splits are not the same, so test set of one config may overlap with samples for another training set. This is usually not a problem but just be aware that you do not train on one config's training set and test with another config's test set.
|
299 |
|
300 |
## Configs
|
301 |
|
302 |
+
### `10k` -- 79 samples -- 10,087 tokens
|
303 |
- ratio_wikipedia: 100.00%
|
304 |
- total_num_tokens: 10,087
|
305 |
- train_num_tokens: 9,205
|
|
|
308 |
- train_num_samples: 78
|
309 |
- test_num_samples: 1
|
310 |
|
311 |
+
### `100k` -- 1,057 samples -- 100,075 tokens
|
312 |
- ratio_wikipedia: 100.00%
|
313 |
- total_num_tokens: 100,075
|
314 |
- train_num_tokens: 98,044
|
|
|
317 |
- train_num_samples: 1,047
|
318 |
- test_num_samples: 10
|
319 |
|
320 |
+
### `1M` -- 10,802 samples -- 1,000,239 tokens
|
321 |
- ratio_wikipedia: 100.00%
|
322 |
- total_num_tokens: 1,000,239
|
323 |
- train_num_tokens: 991,119
|
|
|
326 |
- train_num_samples: 10,694
|
327 |
- test_num_samples: 108
|
328 |
|
329 |
+
### `10M` -- 141,263 samples -- 10,000,022 tokens
|
330 |
- ratio_wikipedia: 100.00%
|
331 |
- total_num_tokens: 10,000,022
|
332 |
- train_num_tokens: 9,874,772
|
|
|
335 |
- train_num_samples: 139,851
|
336 |
- test_num_samples: 1,412
|
337 |
|
338 |
+
### `100M` -- 1,028,484 samples -- 100,000,047 tokens
|
339 |
- ratio_wikipedia: 100.00%
|
340 |
- total_num_tokens: 100,000,047
|
341 |
- train_num_tokens: 99,013,372
|
|
|
344 |
- train_num_samples: 1,018,200
|
345 |
- test_num_samples: 10,284
|
346 |
|
347 |
+
### `1B` -- 5,153,898 samples -- 1,000,000,187 tokens
|
348 |
- ratio_wikipedia: 61.21%
|
349 |
- total_num_tokens: 1,000,000,187
|
350 |
- train_num_tokens: 989,990,190
|
|
|
353 |
- train_num_samples: 5,102,360
|
354 |
- test_num_samples: 51,538
|
355 |
|
356 |
+
### `5B` -- 20,833,009 samples -- 5,000,000,076 tokens
|
357 |
- ratio_wikipedia: 25.35%
|
358 |
- total_num_tokens: 5,000,000,076
|
359 |
- train_num_tokens: 4,984,493,654
|
|
|
362 |
- train_num_samples: 20,769,009
|
363 |
- test_num_samples: 64,000
|
364 |
|
365 |
+
### `10B` -- 40,240,566 samples -- 10,000,000,115 tokens
|
366 |
- ratio_wikipedia: 18.41%
|
367 |
- total_num_tokens: 10,000,000,115
|
368 |
- train_num_tokens: 9,984,156,828
|
|
|
371 |
- train_num_samples: 40,176,566
|
372 |
- test_num_samples: 64,000
|
373 |
|
374 |
+
### `15B` -- 59,648,123 samples -- 15,000,000,154 tokens
|
375 |
- ratio_wikipedia: 15.98%
|
376 |
- total_num_tokens: 15,000,000,154
|
377 |
- train_num_tokens: 14,983,970,518
|
|
|
380 |
- train_num_samples: 59,584,123
|
381 |
- test_num_samples: 64,000
|
382 |
|
383 |
+
### `20B` -- 79,055,679 samples -- 20,000,000,009 tokens
|
384 |
- ratio_wikipedia: 14.75%
|
385 |
- total_num_tokens: 20,000,000,009
|
386 |
- train_num_tokens: 19,983,799,357
|
|
|
389 |
- train_num_samples: 78,991,679
|
390 |
- test_num_samples: 64,000
|
391 |
|
392 |
+
### `25B` -- 98,463,236 samples -- 25,000,000,048 tokens
|
393 |
- ratio_wikipedia: 14.00%
|
394 |
- total_num_tokens: 25,000,000,048
|
395 |
- train_num_tokens: 24,983,765,326
|
|
|
398 |
- train_num_samples: 98,399,236
|
399 |
- test_num_samples: 64,000
|
400 |
|
401 |
+
### `30B` -- 117,870,793 samples -- 30,000,000,087 tokens
|
402 |
- ratio_wikipedia: 13.50%
|
403 |
- total_num_tokens: 30,000,000,087
|
404 |
- train_num_tokens: 29,983,707,932
|