Learning to generate EOS tokens

[vwxyzjn]

@edbeeching and I noticed sometimes the trained SFT models do not learn to stop generations. In other words, the model never learn to generate EOS tokens.

Upon some digging, I noticed this is mainly an issue with the dataset preprocessing. In particular, if we simply pass a dataset like https://huggingface.co/datasets/timdettmers/openassistant-guanaco to the SFTTrainer, the trainer may not postpend the completion with an EOS token.

If we run for item1, item2 in zip(inputs["input_ids"][1], inputs["attention_mask"][1]): print(item1, item2) at https://github.com/huggingface/transformers/blob/91d155ea92da372b319a79dd4eef69533ee15170/src/transformers/trainer.py#L3207, with our SFT example we get

python examples/scripts/sft.py \
    --model_name_or_path="facebook/opt-350m" \
    --report_to="wandb" \
    --learning_rate=1.41e-5 \
    --per_device_train_batch_size=2 \
    --gradient_accumulation_steps=16 \
    --output_dir="sft_openassistant-guanaco" \
    --logging_steps=1 \
    --num_train_epochs=3 \
    --max_steps=-1 \
    --push_to_hub \
    --gradient_checkpointing \
    --dataset_text_field text

Notice how the pad token / eos token corresponds to attention mask = 0.

potential solution

This can be resolved if we add an eos token to the dataset itself. For example,

trl/examples/scripts/minimal/sft.py

Line 57 in dc012ea

"{% for message in messages %}{{' ' + message['content']}}{% endfor %}{{ eos_token }}"

always adds an EOS token to the tokenized dataset, and as a result we get

python examples/scripts/minimal/sft.py \
    --learning_rate 3e-6 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 32 \
    --learning_rate 5e-05 \
    --logging_steps 10 \
    --evaluation_strategy epoch \
    --max_seq_length 1024 \
    --num_train_epochs 5 \
    --output_dir models/minimal/sft

Notice how the first eos token corresponds to attention mask = 1.

[edbeeching]

As I mentioned on out internal slack, we should probably add a line such as:

    if sft_config.packing is False:
        tokenizer.add_eos_token = True

this needs to be removed before saving the model as otherwise generation is broken:

    if sft_config.packing is False:
        # setting this as true breaks generation during evaluation
        tokenizer.add_eos_token = False

I tested these additions in h4 and it resolved many of the issues we saw with models trained with packing=False.

[yananchen1989]

is this an issue when packing=True ? I also do find that the generations from the SFT model are quite wordy.

[derekelewis]

@yananchen1989 I believe the answer is yes for packing=True & packing=False. I'm experiencing lack of predicting EOS on SFTTrainer fine-tuned models w/ using chat templates. Still doing testing, but it doesn't seem to be an issue when not using chat templates and using formatting_func instead.

PEFT also seems to be a contributing factor. No PEFT and EOS is predicted correctly. W/ PEFT and EOS is not correctly predicted.

[vwxyzjn]

Actually I am not even sure if setting the tokenizer.pad_token = tokenizer.eos_token would work. Even if the dataset has an EOS token, what happens is that attention_mask is set to 1, but the label is still set to -100, so the loss on the EOS token is still masked out.

for input_id, attention_mask, label in zip(inputs["input_ids"][0], inputs["attention_mask"][0], inputs["labels"][0]): print(f"{input_id=}, {attention_mask=}, {label=}")

input_id=tensor(15, device='cuda:0'), attention_mask=tensor(1, device='cuda:0'), label=tensor(15, device='cuda:0')
input_id=tensor(0, device='cuda:0'), attention_mask=tensor(1, device='cuda:0'), label=tensor(-100, device='cuda:0')
input_id=tensor(0, device='cuda:0'), attention_mask=tensor(0, device='cuda:0'), label=tensor(-100, device='cuda:0')
input_id=tensor(0, device='cuda:0'), attention_mask=tensor(0, device='cuda:0'), label=tensor(-100, device='cuda:0')
input_id=tensor(0, device='cuda:0'), attention_mask=tensor(0, device='cuda:0'), label=tensor(-100, device='cuda:0')

[yananchen1989]

yes, i agree that no matter packing is set or not, EOS token has not been properly predicted which causes lengthy output.

[vwxyzjn]

@yananchen1989 FYI when packing is set this should not be a problem. See #1646 (comment).

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.