various fixes to the SFT code

model/supervised_finetuning/.gitignore

@@ -0,0 +1 @@
+.cache

model/supervised_finetuning/configs/config.yaml

@@ -13,6 +13,7 @@ defaults:
   logging_steps: 10
   max_grad_norm: 2.0
   save_total_limit: 4
+  fp16: false
   eval_accumulation_steps:
   freeze_layer:
   datasets:

model/supervised_finetuning/custom_datasets/prompt_dialogue.py

@@ -20,7 +20,7 @@ class OAPrivate(Dataset):
         total_prob = reduce(lambda prev, split: prev + split[1], self.splits.items(), 0)
         assert math.isclose(total_prob, 1), "Make sure OAPrivate split ratios add to 1"
 
-        jsonl_file = os.path.join(data_path, self.file)
+        jsonl_file = os.path.join(data_path, file)
 
         with open(jsonl_file, "r", encoding="utf-8") as f:
             lines = f.readlines()

model/supervised_finetuning/trainer.py

@@ -217,7 +217,7 @@ if __name__ == "__main__":
         learning_rate=float(training_conf.learning_rate),
         deepspeed="configs/zero_config.json" if training_conf.deepspeed else None,
         optim=optimizer,
-        fp16=True,
+        fp16=training_conf.fp16,
         local_rank=training_conf.local_rank,
         gradient_checkpointing=training_conf.gradient_checkpointing,
         gradient_accumulation_steps=training_conf.gradient_accumulation_steps,