Unable to reproduce the results from the paper

sunnweiwei / RankGPT

Is ChatGPT Good at Search? LLMs as Re-Ranking Agent [EMNLP 2023 Outstanding Paper Award]

https://arxiv.org/abs/2304.09542

Apache License 2.0

506 stars 46 forks source link

Unable to reproduce the results from the paper #23

Open sina-nabavi opened 5 months ago

sina-nabavi commented 5 months ago

I am unable to reproduce the scores on paper Table 13 using your pre-trained "deberta-10k-rank_net" model. I use the deberta-10k-rank_net function, but the scores that I get are lower.

Moreover, what is the difference between "deberta-v3-large 5K" and "deberta-v3-large 5K label" in the table?

sunnweiwei commented 5 months ago

Hi,

Could you please share the hyperparams used to train the model?

Regarding the figure, the DeBERTa Label 5K (in blue) denotes the model trained using the same hyperparams as the red one but using the MS Marco labels instead of GPT-3.5 predicted labels.