confused with the pooling strategy?

[rxqy]

Hi, I'm confused with the pooling strategy you used here.

For training, you use the avg token

BeLLM/README.md

Line 52 in 9da9269

--pooling_strategy avg \

While for evaluation, you are not specifing any pooling flag here,

BeLLM/README.md

Lines 99 to 105 in 9da9269

	2) evaluate on STS benchmark
	```bash
	BiLLM_START_INDEX=31 CUDA_VISIBLE_DEVICES=0 python eval_sts.py \
	--model_name_or_path NousResearch/Llama-2-7b-hf \
	--lora_name_or_path SeanLee97/bellm-llama-7b-nli \
	--apply_bfloat16 0
	```

so this should be default value [cls], right?

BeLLM/eval_sts.py

Line 57 in 9da9269

parser.add_argument("--pooling_strategy", type=str, default='cls')

As for the paper, you mentioned that you used the representative word as the pivot, so this should be the last non-padding token, right? So I'm wondering which token should I use or does it make no difference in a decoder based model like llama?