IMPORTANT: train on completions only

google · Jul 27, 2024 · d341d61 · d341d61
1 parent c2427ea
commit d341d61
Showing 1 changed file with 6 additions and 1 deletion.
diff --git a/DiarizationLM/unsloth/1_finetune.py b/DiarizationLM/unsloth/1_finetune.py
@@ -8,7 +8,7 @@
 import dataprep
 import torch
 from transformers import TrainingArguments
-from trl import SFTTrainer
+from trl import SFTTrainer, DataCollatorForCompletionOnlyLM
 from unsloth import FastLanguageModel
 from unsloth import is_bfloat16_supported
 
@@ -53,6 +53,10 @@ def run_training() -> None:
   ############################################################################
   # Train the model
   ############################################################################
+  collator = DataCollatorForCompletionOnlyLM(
+      config.PROMPT_SUFFIX.rstrip(),
+      tokenizer=tokenizer)
+
   trainer = SFTTrainer(
       model=model,
       tokenizer=tokenizer,
@@ -61,6 +65,7 @@ def run_training() -> None:
       max_seq_length=config.MAX_SEQ_LENGTH,
       dataset_num_proc=2,
       packing=False,
+      data_collator=collator,
       args=TrainingArguments(
           per_device_train_batch_size=16,
           gradient_accumulation_steps=1,