Merge pull request #30 from winglian/features/python-fixes

backwards support for pre-py3.10, add datasets requirement used in train

finetune.py

@@ -72,14 +72,13 @@ tokenizer.pad_token_id = 0
 if not ft_config.skip:
     # Load Data
     data = None
-    match ft_config.ds_type:
-        case "txt" if not ft_config.skip:
-            #### LLaMA
+    if ft_config.ds_type == "txt" and not ft_config.skip:
+        #### LLaMa
         data = train_data.TrainTxt(ft_config.dataset, ft_config.val_set_size, tokenizer, ft_config.cutoff_len)
-        case "alpaca" if not ft_config.skip:
+    elif ft_config.ds_type == "alpaca" and not ft_config.skip:
         #### Stanford Alpaca-like Data
         data = train_data.TrainSAD(ft_config.dataset, ft_config.val_set_size, tokenizer, ft_config.cutoff_len)
-        case _:
+    else:
         raise NotImplementedError("ERROR: Unknown dataset format")
     data.prepare_data()
     ####

requirements.txt

@@ -1,6 +1,8 @@
 torch
 accelerate
 bitsandbytes
+datasets
+sentencepiece
 git+https://github.com/huggingface/transformers.git
 git+https://github.com/sterlind/GPTQ-for-LLaMa.git@lora_4bit
 git+https://github.com/sterlind/peft.git