max_length decides what the model learns from.
Three Larkspur Cloud support examples in the lesson 45.4 template, tokenized with the Pythia tokenizer the notebooks use. Move max_length, pick the side to cut and the padding, and read the batch the Trainer receives.
- Answer tokens cut
- Question tokens cut
- Padding share of the batch
- Rows that end with the end token
The batch the Trainer receives
EleutherAI/pythia-70m tokenizer
What the tensors hold
| Example | Tokens | Kept | Padding | Mask ones | Row width |
|---|
The notebook sets pad_token = eos_token, so every padding position holds id 0, the same id as <|endoftext|>. labels is a copy of input_ids, as in lesson 45.4. Checked against transformers 5.17 on every setting of the four controls.