set labels and fix datasets block

requirements.txt

@@ -12,7 +12,7 @@ liger-kernel==0.4.2
 
 packaging==23.2
 peft==0.14.0
-transformers>=4.46.3
+transformers==4.47.0
 tokenizers>=0.20.1
 accelerate==1.2.0
 datasets==3.1.0

src/axolotl/prompt_strategies/pretrain.py

@@ -41,6 +41,7 @@ class PretrainTokenizationStrategy(PromptTokenizingStrategy):
             seq + [self.tokenizer.eos_token_id] for seq in res["input_ids"]
         ]
         res["attention_mask"] = [seq + [1] for seq in res["attention_mask"]]
+        res["labels"] = res["input_ids"].copy()
 
         return res
 
@@ -49,12 +50,16 @@ class PretrainTokenizationStrategy(PromptTokenizingStrategy):
 
 
 def load(tokenizer, cfg):
+    if cfg.pretraining_dataset:
+        cfg_ds = cfg.pretraining_dataset
+    else:
+        cfg_ds = cfg.datasets
     strat = PretrainTokenizationStrategy(
         PretrainTokenizer(),
         tokenizer,
         cfg.train_on_inputs,
         cfg.sequence_len,
-        text_column=cfg.pretraining_dataset[0]["text_column"] or "text",
+        text_column=cfg_ds[0]["text_column"] or "text",
         max_length=cfg.sequence_len * 64,
     )
     return strat