diff --git a/Ordinal/build_Amazon_datasets.py b/Ordinal/build_Amazon_datasets.py
index 533f7e4..cf29eb9 100644
--- a/Ordinal/build_Amazon_datasets.py
+++ b/Ordinal/build_Amazon_datasets.py
@@ -1,14 +1,17 @@
 import gzip
 import quapy as qp
+from Ordinal.utils import load_simple_sample_raw
 from quapy.data import LabelledCollection
 import quapy.functional as F
 import os
 from os.path import join
 from pathlib import Path
+import numpy as np
 
 
 datadir = '/mnt/1T/Datasets/Amazon/reviews'
 outdir  = './data/'
+real_prev_path = './data/Books-real-prevalence-by-product_votes1_reviews100.csv'
 domain = 'Books'
 seed = 7
 
@@ -63,7 +66,6 @@ def gen_samples_APP(pool: LabelledCollection, nsamples, sample_size, outdir, pre
             write_txt_sample(sample, join(outdir, f'{i}.txt'))
             prevfile.write(f'{i},' + ','.join(f'{p:.3f}' for p in sample.prevalence()) + '\n')
 
-
 def gen_samples_NPP(pool: LabelledCollection, nsamples, sample_size, outdir, prevpath):
     os.makedirs(outdir, exist_ok=True)
     with open(prevpath, 'wt') as prevfile:
@@ -73,37 +75,69 @@ def gen_samples_NPP(pool: LabelledCollection, nsamples, sample_size, outdir, pre
             prevfile.write(f'{i},' + ','.join(f'{p:.3f}' for p in sample.prevalence()) + '\n')
 
 
+def gen_samples_real_prevalences(real_prevalences, pool: LabelledCollection, sample_size, outdir, prevpath_out):
+    os.makedirs(outdir, exist_ok=True)
+    with open(prevpath_out, 'wt') as prevfile:
+        prevfile.write('id,' + ','.join(f'{c}' for c in pool.classes_) + '\n')
+        for i, prev in enumerate(real_prevalences):
+            sample = pool.sampling(sample_size, *prev[:-1])
+            write_txt_sample(sample, join(outdir, f'{i}.txt'))
+            prevfile.write(f'{i},' + ','.join(f'{p:.3f}' for p in sample.prevalence()) + '\n')
 
-fullpath = join(datadir,domain)+'.txt.gz'
 
-data = LabelledCollection.load(fullpath, from_gz_text)
-print(len(data))
-print(data.classes_)
-print(data.prevalence())
+# fullpath = join(datadir,domain)+'.txt.gz'
+#
+# data = LabelledCollection.load(fullpath, from_gz_text)
+# print(len(data))
+# print(data.classes_)
+# print(data.prevalence())
 
 with qp.util.temp_seed(seed):
-    train, rest = data.split_stratified(train_prop=tr_size)
-
-    devel, test = rest.split_stratified(train_prop=0.5)
-    print(len(train))
-    print(len(devel))
-    print(len(test))
-
+    # train, rest = data.split_stratified(train_prop=tr_size)
+    #
+    # devel, test = rest.split_stratified(train_prop=0.5)
+    # print(len(train))
+    # print(len(devel))
+    # print(len(test))
+    #
     domaindir = join(outdir, domain)
 
-    write_txt_sample(train, join(domaindir, 'training_data.txt'))
-    write_txt_sample(devel, join(domaindir, 'development_data.txt'))
-    write_txt_sample(test, join(domaindir, 'test_data.txt'))
+    # write_txt_sample(train, join(domaindir, 'training_data.txt'))
+    # write_txt_sample(devel, join(domaindir, 'development_data.txt'))
+    # write_txt_sample(test, join(domaindir, 'test_data.txt'))
 
-    gen_samples_APP(devel, nsamples=nval, sample_size=val_size, outdir=join(domaindir, 'app', 'dev_samples'),
-                    prevpath=join(domaindir, 'app', 'dev_prevalences.txt'))
-    gen_samples_APP(test, nsamples=nte, sample_size=te_size, outdir=join(domaindir, 'app', 'test_samples'),
-                    prevpath=join(domaindir, 'app', 'test_prevalences.txt'))
+    # this part is to be used when the partitions have already been created, in order to avoid re-generating them
+    train = load_simple_sample_raw(domaindir, 'training_data')
+    devel = load_simple_sample_raw(domaindir, 'development_data')
+    test = load_simple_sample_raw(domaindir, 'test_data')
 
-    gen_samples_NPP(devel, nsamples=nval, sample_size=val_size, outdir=join(domaindir, 'npp', 'dev_samples'),
-                    prevpath=join(domaindir, 'npp', 'dev_prevalences.txt'))
-    gen_samples_NPP(test, nsamples=nte, sample_size=te_size, outdir=join(domaindir, 'npp', 'test_samples'),
-                    prevpath=join(domaindir, 'npp', 'test_prevalences.txt'))
+    # gen_samples_APP(devel, nsamples=nval, sample_size=val_size, outdir=join(domaindir, 'app', 'dev_samples'),
+    #                 prevpath=join(domaindir, 'app', 'dev_prevalences.txt'))
+    # gen_samples_APP(test, nsamples=nte, sample_size=te_size, outdir=join(domaindir, 'app', 'test_samples'),
+    #                 prevpath=join(domaindir, 'app', 'test_prevalences.txt'))
+
+    # gen_samples_NPP(devel, nsamples=nval, sample_size=val_size, outdir=join(domaindir, 'npp', 'dev_samples'),
+    #                 prevpath=join(domaindir, 'npp', 'dev_prevalences.txt'))
+    # gen_samples_NPP(test, nsamples=nte, sample_size=te_size, outdir=join(domaindir, 'npp', 'test_samples'),
+    #                 prevpath=join(domaindir, 'npp', 'test_prevalences.txt'))
+
+
+    # this part generates samples based on real prevalences (in this case, prevalences of sets of books reviews
+    # groupped by product). It loads the real prevalences (computed elsewhere), and randomly extract 5000 for test
+    # and 1000 for val (disjoint). Then realize the samplings
+
+    assert os.path.exists(real_prev_path), f'real prevalence file does not seem to exist...'
+    real_prevalences = np.genfromtxt(real_prev_path, delimiter='\t')
+
+    nrows = real_prevalences.shape[0]
+    rand_sel = np.random.permutation(nrows)
+    real_prevalences_val = real_prevalences[rand_sel[:nval]]
+    real_prevalences_te  = real_prevalences[rand_sel[nval:nval+nte]]
+
+    gen_samples_real_prevalences(real_prevalences_val, devel, sample_size=val_size, outdir=join(domaindir, 'real', 'dev_samples'),
+                    prevpath_out=join(domaindir, 'real', 'dev_prevalences.txt'))
+    gen_samples_real_prevalences(real_prevalences_te, test, sample_size=te_size, outdir=join(domaindir, 'real', 'test_samples'),
+                    prevpath_out=join(domaindir, 'real', 'test_prevalences.txt'))
 
 
 
diff --git a/Ordinal/generate_bert_vectors_npytxt.py b/Ordinal/generate_bert_vectors_npytxt.py
index f58d5ae..2e83ae4 100644
--- a/Ordinal/generate_bert_vectors_npytxt.py
+++ b/Ordinal/generate_bert_vectors_npytxt.py
@@ -89,6 +89,7 @@ def get_best_checkpoint(checkpointdir):
         steps.append(step)
     assert len(steps) <= 2, 'unexpected number of steps, only two where expected (the best one and the last one)'
     choosen = f'{checkpointdir}/checkpoint-{min(steps)}'
+    print(f'choosen checkpoint is {choosen}')
     return choosen
 
 
@@ -97,7 +98,7 @@ if __name__ == '__main__':
     assert torch.cuda.is_available(), 'cuda is not available'
 
     #checkpoint='roberta-base-val-finetuned'
-    #generation_mode = 'posteriors'
+    #generation_mode = 'ave'
 
     n_args = len(sys.argv)
     assert n_args==3, 'wrong arguments, expected: <checkpoint> <generation-mode>\n' \
@@ -114,7 +115,7 @@ if __name__ == '__main__':
 
     datapath = './data'
     domain = 'Books'
-    protocols = ['app']  # ['app', 'npp']
+    protocols = ['real']  # ['app', 'npp']
 
     assert generation_mode in ['last', 'average', 'posteriors'], 'unknown generation_model'
     outname = domain + f'-{checkpoint}-{generation_mode}'