fix: Move deprecated positional arguments from SFTTrainer to SFTConfig (#399)

* fix: set legacy behavior to false, enable new behavior

Signed-off-by: Will Johnson <[email protected]>

* fix: Resolve push_to_hub_token warning

Signed-off-by: Will Johnson <[email protected]>

* fix: Remove max_seq_length and dataset_text_field from SFTTrainer

Signed-off-by: Will Johnson <[email protected]>

* fmt

Signed-off-by: Will Johnson <[email protected]>

* fix: Resolve tokenizer.padding_side warning

Signed-off-by: Will Johnson <[email protected]>

* nit: restructure warning fixes

Signed-off-by: Will Johnson <[email protected]>

* fix: Add packing directly to SFTConfig

Signed-off-by: Will Johnson <[email protected]>

* fmt

Signed-off-by: Will Johnson <[email protected]>

* Removed dataset_kwargs from SFTTrainer

Removed the argument dataset_kwargs from the the invocation of SFTTRainer() because it will be deprecated in V1.0.0. Instead, dataset_kwargs have been added as a key to the training_args variable.

Following the example provided by HF found here: https://huggingface.co/docs/trl/en/sft_trainer#training-the-vision-language-model

Signed-off-by: Luka Dojcinovic <[email protected]>

* fix: Added max_seq_length back to SFTConfig()

Signed-off-by: Luka Dojcinovic <[email protected]>

* Removed legacy and padding_side args

Removed these args as they were based on changes from @willmj that haven't been approved yet

Signed-off-by: Luka Dojcinovic <[email protected]>

* Moved all args to additional_args

Following @kmehant suggestion.

Signed-off-by: Luka Dojcinovic <[email protected]>

* Removed packing and max_seq_length

Removed packing and max_seq_length variables from additional_args

Signed-off-by: Luka Dojcinovic <[email protected]>

* Removed check is_pretokenized_dataset

Co-authored-by: Mehant Kammakomati <[email protected]>
Signed-off-by: Luka-D <[email protected]>

* Removed max_seq_length from additional_args

Signed-off-by: Luka Dojcinovic <[email protected]>

* Removed error.log

Signed-off-by: Luka Dojcinovic <[email protected]>

* fix: move packing to SFTConfig as well

Co-authored-by: Luka-D <[email protected]>
Signed-off-by: Mehant Kammakomati <[email protected]>

---------

Signed-off-by: Will Johnson <[email protected]>
Signed-off-by: Luka Dojcinovic <[email protected]>
Signed-off-by: Luka-D <[email protected]>
Signed-off-by: Mehant Kammakomati <[email protected]>
Co-authored-by: Will Johnson <[email protected]>
Co-authored-by: Mehant Kammakomati <[email protected]>
Co-authored-by: Mehant Kammakomati <[email protected]>

Author: 3 people

tuning/sft_trainer.py

@@ -318,27 +318,29 @@ def train(
     # this validation, we just drop the things that aren't part of the SFT Config and build one
     # from our object directly. In the future, we should consider renaming this class and / or
     # not adding things that are not directly used by the trainer instance to it.
+
     transformer_train_arg_fields = [x.name for x in dataclasses.fields(SFTConfig)]
     transformer_kwargs = {
         k: v
         for k, v in train_args.to_dict().items()
         if k in transformer_train_arg_fields
     }
-    training_args = SFTConfig(**transformer_kwargs)
+
+    additional_args = {
+        "dataset_text_field": dataset_text_field,
+        "dataset_kwargs": dataset_kwargs,
+    }
+    training_args = SFTConfig(**transformer_kwargs, **additional_args)
 
     trainer = SFTTrainer(
         model=model,
         tokenizer=tokenizer,
         train_dataset=formatted_train_dataset,
         eval_dataset=formatted_validation_dataset,
-        packing=train_args.packing,
         data_collator=data_collator,
-        dataset_text_field=dataset_text_field,
         args=training_args,
-        max_seq_length=max_seq_length,
         callbacks=trainer_callbacks,
         peft_config=peft_config,
-        dataset_kwargs=dataset_kwargs,
     )
 
     # We track additional metrics and experiment metadata after trainer object creation