yyDing1 / GNER

Rethinking Negative Instances for Generative Named Entity Recognition
Apache License 2.0
45 stars 2 forks source link

Unable to reproduce the results of the paper. #5

Closed wyclike closed 1 month ago

wyclike commented 1 month ago

Hello! Can you provide the inference results in a jsonl file on the CONLL2003 test set? I am unable to replicate the results you achieved on this dataset, and there is a difference of 20 points.

wyclike commented 1 month ago

i use your code and the model of "GNER-LLaMA-7B'

wyclike commented 1 month ago

could u give the ‘llama7b-supervised.jsonl’?

yyDing1 commented 1 month ago

Thank you for your attention!

The GNER-LLaMA-7B is the model after task adaptation (Stage 1). The CoNLL2003 Test results reported in the paper are based on the GNER-LLaMA-7B after supervised fine-tuning (Stage 2).

We provide the generation results for quick reproductions here, where xxx-stage2 refers to the supervised fine-tuning results. For Llama-7B, you can use the publicly available dataset and our training code to reproduce the results.

wyclike commented 1 month ago

Yeah, I knew. But there is no ‘llama7b-supervised.jsonl’ in the Google Drive folder; there are only task adaptation results.

wyclike commented 1 month ago

or could you please provide the post-supervised checkpoint for the llama7b ? Thank you!

yyDing1 commented 1 month ago

predictions.jsonl.zip

wyclike commented 1 month ago

Thanks!