🪁 (#129)

* Add Gemma 7B recipe

* Use Gemma template

* Make it work for dolly lol

* Enable cahce

* Clean up

* DPO to the max

* DPO, DPO, DPO

* Add openhermes

* Add custom configs

* Add kwargs

* Fix config

* Bump deps

* Move old recipes

* Add doc

* Add norte

* Renable cache

* Nuke

* Clean

* Apply suggestions from code review

Co-authored-by: Alvaro Bartolome <alvaro@argilla.io>

* Fix isort

* Update README.md

* Update config_full.yaml

---------

Co-authored-by: Alvaro Bartolome <alvaro@argilla.io>
Co-authored-by: Philipp Schmid <32632186+philschmid@users.noreply.github.com>

src/alignment/configs.py

@@ -136,6 +136,14 @@ class ModelArguments:
             "choices": ["auto", "bfloat16", "float16", "float32"],
         },
     )
+    tokenizer_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": (
+                "The path to the tokenizer. Useful if you want to use a different tokenizer to the one stored in `model_name_or_path`."
+            )
+        },
+    )
     trust_remote_code: bool = field(default=False, metadata={"help": "Trust remote code when loading a model."})
     use_flash_attention_2: bool = field(
         default=False,
@@ -220,6 +228,9 @@ class SFTConfig(transformers.TrainingArguments):
     Arguments related to the training process itself. For all parameters, see: https://huggingface.co/docs/transformers/v4.26.1/en/main_classes/trainer#transformers.TrainingArguments
     """
 
+    dataset_kwargs: Optional[Dict[str, Any]] = field(
+        default=None, metadata={"help": "Dataset kwargs for the SFTTrainer"}
+    )
     max_seq_length: Optional[int] = field(
         default=None,
         metadata={"help": ("Used by TRL for reward model training, which tries to read this parameter in init.")},

src/alignment/data.py

@@ -34,7 +34,7 @@ def maybe_insert_system_message(messages, tokenizer):
         chat_template = tokenizer.default_chat_template
 
     # confirm the jinja template refers to a system message before inserting
-    if "system" in chat_template:
+    if "system" in chat_template or "<|im_start|>" in chat_template:
         messages.insert(0, {"role": "system", "content": ""})
 
 

src/alignment/model_utils.py

@@ -65,7 +65,9 @@ def get_quantization_config(model_args: ModelArguments) -> BitsAndBytesConfig |
 def get_tokenizer(model_args: ModelArguments, data_args: DataArguments) -> PreTrainedTokenizer:
     """Get the tokenizer for the model."""
     tokenizer = AutoTokenizer.from_pretrained(
-        model_args.model_name_or_path,
+        model_args.model_name_or_path
+        if model_args.tokenizer_name_or_path is None
+        else model_args.tokenizer_name_or_path,
         revision=model_args.model_revision,
     )
     if tokenizer.pad_token_id is None: