Fix IDs shape mismatch in SFT for VLMs with text-only

[albertvillanova]

Fix IDs shape mismatch in SFT for VLMs with text-only.

Fix #5334.

This PR addresses a regression issue when training vision-language models (VLMs) with text-only datasets, ensuring compatibility between data preprocessing and model expectations. The main focus is on fixing how input IDs are handled for VLMs and adding a regression test to prevent future breakage.

Changes

Bug fix for VLM text-only input handling:

• Fixed an inconsistency in tokenize_fn where VLM processors returned input IDs as a list of lists (e.g., [[1, 2, 3]]) instead of a flat list (e.g., [1, 2, 3]). The function now unwraps the extra list level to prevent downstream shape errors in models expecting 3-D position IDs.

Testing improvements:

• Added a regression test in test_sft_trainer.py to verify that training a VLM with a text-only dataset works correctly and does not produce shape errors.

---

Note

Medium Risk
Touches SFT dataset tokenization for standard (non-conversational) examples; a small shape-normalization change could affect any processor that returns nested input_ids, but it is guarded by a targeted regression test.

Overview
Fixes a regression when training vision-language models on text-only standard datasets by normalizing input_ids returned from VLM processing_class calls (unwrapping [[...]] to [...]) to match the LLM code path and avoid downstream shape/position-id errors.

Adds a regression case to test_train_vlm_text_only_data to include standard_language_modeling in the parameterized dataset configs, ensuring VLM text-only training remains supported.

^{Written by Cursor Bugbot for commit 7b1acb1. This will update automatically on new commits. Configure here.}

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[qgallouedec]

For the records: we do not have the same issue in DPO (text-only data is properly supported), neither in RewardTrainer which doesn't support VLM