Model training not yielding expected results

[Sayil8]

Describe the bug
After training a model with ATEPCTrainer using Spanish data, including specific reviews I added, the model fails to correctly identify aspects/sentiments—even on the exact same reviews it was trained on. For example, the model predicts aspects/sentiments that do not match the training data.

---

To Reproduce
Steps to reproduce the behavior:

1. Use the following dataset for training:
   https://github.com/yangheng95/ABSADatasets/blob/v2.0/datasets/atepc_datasets/120.SemEval2016Task5/127.spanish/restaurants_train_spanish.xml.dat.atepc
   (plus custom Spanish reviews, see below)

2. Add this example to the training data:

   Muy O -100
   amables B-ASP Positive
   todos O -100
   ! O -100
   Betty O -100
   súper O -100
   amable B-ASP Positive
   y O -100
   simpática O -100
   

3. Train the model using the following code:

   import os
   from pyabsa import AspectTermExtraction as ATEPC
   
   if __name__ == '__main__':
       dataset_name = "MySpanishRestaurant"
       config = ATEPC.ATEPCConfigManager.get_atepc_config_multilingual()
       config.model = ATEPC.ATEPCModelList.FAST_LCF_ATEPC
       config.pretrained_bert = "BSC-LT/roberta-base-bne"
       config.num_epoch = 20
       config.evaluate_begin = 0
       config.max_seq_len = 128
       config.log_step = 10
       config.learning_rate = 2e-5
       config.batch_size = 16
       config.patience = 3
       config.device = "auto"
       config.log_step = -1
       config.l2reg = 1e-8
       config.seed = 42
       config.load_aug = True
   
       print("Starting model training...")
       trainer = ATEPC.ATEPCTrainer(
           config=config,
           dataset=dataset_name
       )
       trained_model_path = trainer.inference_model
       print(f"Training complete! Model saved at: {trained_model_path}")
       if os.path.exists(trained_model_path):
           print("Verification successful: The model file exists at the specified path.")
       else:
           print("Verification FAILED: The model file does NOT exist at the specified path.")

4. After training, test the model on the same review:

   Muy amables todos ! Betty súper amable y simpática
   

---

Actual behavior
The model outputs:

Aspects/Sentiments Found: [('todos', 'Positive'), ('y', 'Positive')]

This does not match the expected aspects/sentiments from the training data.

---

Expected behavior
The model should correctly identify the aspects and sentiments as provided in the training data, especially when evaluating on the exact same examples.

---

Environment:

• PyABSA version: 2.4.2
• Transformers version: 4.56.0
• Torch version: 2.8.0+cudaNone
• Device: Unknown

---

Additional context

• I have more examples with similar issues and can share them if needed.
• Training logs

---

Screenshots
N/A

[Sayil8]

@yangheng95 is this something you could look into?

[yangheng95]

Sorry for the late reply. So, in this context, have the positions been identified in errors? Does this happen to all the datasets?

[Sayil8]

No worries, thanks for the reply and taking a look.

Yes. The model is identifying the wrong positions/words as aspects.

So far, I have only tested this with the Spanish dataset (restaurants_train_spanish.xml.dat.atepc) combined with my custom examples. I have not tried another dataset yet.

[yangheng95]

Sorry for this issue. I am very busy in other projects, so may you help me confirm if the error happens to different models and datasets? That would be helpful in the problem location and repair.