Point all references at NAMAA-Space; document the fixed tokenizer and a version-robust loader

train_nilechat_qlora.py

@@ -9,12 +9,12 @@ as the organisers' official baseline) for English -> dialectal Arabic dialogue t
 One script, three variants - the ablation this pair of repos exists to document:
 
     --variant noctx    the prompt's history block is always "No previous turns"
-                       -> FatimahEmadEldin/alexandriax-nilechat-lora
+                       -> NAMAA-Space/alexandriax-nilechat-lora
                           dev 23.54 spBLEU / 39.68 chrF++   (BEST of the three)
     --variant ctx      the model's own previous turns are fed back as history
                        -> not released;  dev 22.87 / 39.11
     --variant ctx_aux  ctx + 348,787 back-translated pairs
-                       -> FatimahEmadEldin/alexandriax-nilechat-ctx-aux
+                       -> NAMAA-Space/alexandriax-nilechat-ctx-aux
                           dev 22.77 / 38.71
 
 "dev" is the official 12,250-turn AlexandriaX development set, country-macro spBLEU
@@ -57,7 +57,7 @@ from trl import SFTConfig, SFTTrainer
 # --------------------------------------------------------------------------------------
 MODEL_NAME = "UBC-NLP/NileChat-3B-Base"
 DATASET = "UBC-NLP/alexandria"
-AUX_DATASET = "FatimahEmadEldin/alexandria-backtranslated-pairs"   # 348,787 synthetic pairs
+AUX_DATASET = "NAMAA-Space/alexandria-backtranslated-pairs"   # 348,787 synthetic pairs
 
 MAX_SEQ_LENGTH = 1024
 MAX_NEW_TOKENS = 128