TypeError when saving predictions with HuggingFace Dataset

[pyxnpyx]

Reminder

• I have read the above rules and searched the existing issues.

System Info

[Bug] TypeError when saving predictions with HuggingFace Dataset

Description

When running evaluation with a local dataset, saving predictions fails with:

TypeError: argument 'ids': 'list' object cannot be interpreted as an integer

Line 179 in ‘LLaMA-Factory/src/llamafactory/train/sft/trainer.py`:

decoded_inputs = self.processing_class.batch_decode(dataset["input_ids"], skip_special_tokens=False)

Root Cause

It seems that with local datasets, LLaMA-Factory converts them to HuggingFace Dataset format internally. So dataset["input_ids"] returns a Column object instead of a Python list, and batch_decode() expects a list.

Solution

Convert Column to list:

input_ids_list = dataset["input_ids"].to_pylist()
decoded_inputs = self.processing_class.batch_decode(input_ids_list, skip_special_tokens=False)

Question

Is this a bug? If it's a bug, I'd be happy to submit a fix.
Thanks!

Reproduction

Reproduce

1. Use a local JSON dataset
2. Run evaluation with
   "stage: sft
   do_predict: true
   finetuning_type: lora
   adapter_name_or_path:...
   "
3. Error occurs when saving predictions

Others

No response

[majiayu000]

Hi, I've been looking into this and traced the root cause.

Proposed approach: In src/llamafactory/train/sft/trainer.py line 179, change dataset['input_ids'] to dataset['input_ids'].to_pylist() so the HuggingFace Dataset Column object is converted to a Python list before passing to batch_decode(). Also check if similar patterns exist nearby (e.g., dataset['labels'] access) that may need the same fix.

I'll open a draft PR shortly. Happy to adjust based on your preference.

[majiayu000]

Hi, I've been looking into this and traced the root cause.

Proposed approach: In src/llamafactory/train/sft/trainer.py line 179, convert the HuggingFace Dataset Column object to a Python list before passing to batch_decode(). Change dataset['input_ids'] to list(dataset['input_ids']) or .to_pylist(). Check nearby code (e.g., dataset['labels']) for the same pattern. Add a defensive conversion to avoid the TypeError.

I'll open a draft PR shortly. Happy to adjust based on your preference.

[majiayu000]

Hi, I've been looking into this and traced the root cause.

Proposed approach: In src/llamafactory/train/sft/trainer.py line 179, convert dataset['input_ids'] Column object to a Python list before passing to batch_decode(). Use list() or .to_pylist(). Check nearby lines (e.g., dataset['labels']) for the same pattern. Minimal one-file change, no tests expected for this kind of fix in this repo.

I'll open a draft PR shortly. Happy to adjust based on your preference.