Docs: align Unit 2 DPOConfig & Unit 3 VLM SFT format with current TRL (1.5)

[ColebyPearson]

Two small but reproducible API-drift errors learners hit when running units/en/ verbatim against TRL 1.5+.

1. units/en/unit2/2.md — drop max_prompt_length from the DPOConfig example

TRL 1.5 removed the max_prompt_length argument. The current snippet (lines 86–94) raises immediately:

TypeError: DPOConfig.__init__() got an unexpected keyword argument 'max_prompt_length'

max_length alone caps the combined prompt+completion sequence, so max_prompt_length is redundant. Removed the line and updated the comment on max_length to make its new role explicit.

2. units/en/unit3/4.md — format_data() uses typed-parts message content that TRL's SFT collator rejects

The current format_data() builds messages like:

{"role": "user", "content": [{"type": "image", ...}, {"type": "text", ...}]}

TRL's SFT collator (trl.data_utils.prepare_multimodal_messages) counts <image> placeholders in string content and auto-injects them from the images field. The typed-parts form is invisible to it, so training raises:

ValueError: Number of images provided (1) does not match number of image placeholders (0).

Switching to plain-string content lets TRL inject the placeholder itself. I added a short comment in the snippet explaining the why, so the next learner who runs into a related collator error has something greppable.

The earlier apply_chat_template usage in the same file (Exercise 1, lines ~148–152) is left untouched — typed-parts work fine for the processor's chat-template path; only the SFTTrainer collator is brittle.

Verification

• Diff is intentionally surgical: 3 insertions / 35 deletions across both files, no whitespace/line-ending noise.
• Hit both while running the unit verbatim from a clean Python 3.13 + TRL 1.5 + PEFT 0.18 install (notes in VoicesColeby/HFsmolcourse if useful).
• Tested by re-running the Unit 2 + Unit 3 scripts against current TRL — both now train end-to-end and push to the Hub.

🤖 Generated with Claude Code