diff --git a/KoozDawa/tweet.py b/KoozDawa/tweet.py
index 0799730..1e43264 100644
--- a/KoozDawa/tweet.py
+++ b/KoozDawa/tweet.py
@@ -19,6 +19,8 @@ def tweet():
     # des glaçons pour les yeux brisées
     # je suis pas juste un verbe que t'observe
 
+    # si tu sais rien pas d'âme de la vie
+
     Tweeper("KoozDawa").tweet("tassepés en panel")
 
 
diff --git a/LeBoulbiNet/boulbi.py b/LeBoulbiNet/boulbi.py
index ee91b16..d9e4e2a 100644
--- a/LeBoulbiNet/boulbi.py
+++ b/LeBoulbiNet/boulbi.py
@@ -1,65 +1,50 @@
+from datetime import datetime
+
 from keras.callbacks import ModelCheckpoint, EarlyStopping
 
 from glossolalia.loader import load_seeds, load_text
-from glossolalia.lstm import generate_padded_sequences, create_model, generate_text
-from glossolalia.tokens import PoemTokenizer
+from glossolalia.lstm import LisSansTaMaman
 
 
-def main():
+def train():
     # should_train = True
-    # model_file = "../models/dawa_lstm_%i.hd5" % nb_epoch
     nb_words = 20
     nb_epoch = 50
     nb_layers = 64
     dropout = .2
-    tokenizer = PoemTokenizer()
+    # TODO finetune layers/dropout
+    validation_split = 0.2
+    lstm = LisSansTaMaman(nb_layers, dropout, validation_split, debug=True)
+    filename_model = "../models/boulbi/boulbi_lstm%i-d%.1f-{epoch:02d}_%i-{accuracy:.4f}.hdf5" % (
+        nb_layers, dropout, nb_epoch)
+    filename_output = "./output/boulbi_%i-d%.1f_%s.txt" % (
+        nb_layers, dropout, datetime.now().strftime("%y%m%d_%H%M"))
+    callbacks_list = [
+        ModelCheckpoint(filename_model, monitor='val_accuracy', period=10, save_best_only=True),
+        EarlyStopping(monitor='val_accuracy', patience=5)]
 
-    # if should_train:
     corpus = load_text()
     print("Corpus:", corpus[:10])
+    lstm.create_model(corpus)
+    with open(filename_output, "a+") as f:
+        for i in range(0, nb_epoch, 10):
+            lstm.fit(epochs=min(i + 10, nb_epoch), initial_epoch=i,
+                     callbacks=callbacks_list,
+                     validation_split=validation_split)
 
-    inp_sequences, total_words = tokenizer.get_sequence_of_tokens(corpus)
-    predictors, label, max_sequence_len = generate_padded_sequences(inp_sequences, total_words)
-    model = create_model(max_sequence_len, total_words, layers=nb_layers, dropout=dropout)
-    model.summary()
-
-    file_path = "../models/boulbi/boulbi_lstm%i-d%.1f-{epoch:02d}_%i-{accuracy:.4f}.hdf5" % (nb_layers, dropout, nb_epoch)
-    checkpoint = ModelCheckpoint(file_path, monitor='accuracy', period=10, save_best_only=True)
-    # print_callback = LambdaCallback(on_epoch_end=on_epoch_end)
-    early_stopping = EarlyStopping(monitor='accuracy', patience=5)
-    callbacks_list = [checkpoint, early_stopping]
-
-    for i in range(0, nb_epoch, 10):
-        model.fit(predictors, label, initial_epoch=i, epochs=min(i + 10, nb_epoch), verbose=2, callbacks=callbacks_list)
-        for seed in ["", "Je", "Tu", "Le", "La", "Les", "Un", "On", "Nous"]:
-            print(generate_text(model, tokenizer, seed, nb_words, max_sequence_len))
-
-    # model.save(model_file)
-    # else: # FIXME: Load and predict, maybe reuse checkpoints?
-    # model = load_model(model_file)
+            print(lstm.predict_seeds(nb_words))
 
-    for i, seed in enumerate(load_seeds(corpus, 5)):
-        output = generate_text(model, tokenizer, seed, nb_words, max_sequence_len)
-        print("%i %s -> %s" % (i, seed, output))
+        for i, seed in enumerate(load_seeds(corpus, 5)):
+            output = lstm.predict(seed, nb_words)
+            print("%i %s -> %s" % (i, seed, output))
+            f.writelines(output)
 
-    with open("./output/boulbi.txt", "a+") as f:
         while True:
             input_text = input("> ")
-            text = generate_text(model, tokenizer, input_text, nb_words, max_sequence_len)
-
+            text = lstm.predict(input_text, nb_words)
             print(text)
             f.writelines("%s\n" % text)
 
 
-def debug_unrandomize():
-    from numpy.random import seed
-    from tensorflow_core.python.framework.random_seed import set_random_seed
-
-    # set seeds for reproducibility
-    set_random_seed(2)
-    seed(1)
-
-
 if __name__ == '__main__':
-    debug_unrandomize()
-    main()
+    train()
diff --git a/glossolalia/lstm.py b/glossolalia/lstm.py
index 51dec94..a695627 100644
--- a/glossolalia/lstm.py
+++ b/glossolalia/lstm.py
@@ -56,7 +56,7 @@ class LisSansTaMaman(object):
                               validation_split=validation_split,
                               epochs=epochs, initial_epoch=initial_epoch)
 
-    def predict_seeds(self, seeds: List[str] = None, nb_words=None):
+    def predict_seeds(self, nb_words=None, seeds: List[str] = None):
         if seeds is None:
             seeds = ["", "Je", "Tu", "Le", "La", "Les", "Un", "On", "Nous"]
         return [self.predict(seed, nb_words) for seed in seeds]