cgq15 / dect Goto Github PK

Hi～ @cgq15 ,
I’ve been closely studying your recent work and am greatly impressed by the results you’ve achieved. It’s been a great source of inspiration for my own research.
I’m currently reproducing the results in Table 5. However, I’m facing some challenges with the NLI datasets (i.e., RTE, SNLI, and MNLI), my results of these datasets are quite a bit lower than the results reported (e.g., getting only 38.6 and 36.9 for SNLI with T5-base and T5-large).

For your reference, here is the script I have been using for run_dect.py:

python src/run_dect.py \
        --model t5 \
        --size base \
        --type lm \
        --model_name_or_path t5-base \
        --shot 16 \
        --dataset snli \
        --lr=0.01

I set the argument 'type' to 'lm', so the template for NLI dataset is
{"placeholder": "text_a"} {"placeholder": "text_b"} Does the first sentence entail the second sentence? {"mask"}

I’m reaching out to kindly ask if you could share any specific templates or settings you used for the T5 model, particularly for the NLI task datasets. Any insight or details you could provide about your experimental setup would be immensely helpful.

Thank you so much for your time.