train_with_paraphrasing
Collection
[long-context models trained with "original text paraphrasing" dataset](https://github.com/yuyijiong/train_with_paraphrasing)
•
4 items
•
Updated
Qwen2-7B-Instruct模型在 Long-Instruction-with-Paraphrasing数据集上微调 1 epoch,提升了 long-context 能力
long-context 能力得到提升
dataset | Qwen2-7B-Instruct | Qwen2-7b-Instruct-paraph |
---|---|---|
hotpotqa | 42.79 | 49.46 |
dureader | 24.28 | 33.94 |
multifieldqa_en | 46.17 | 49.65 |
multifieldqa_zh | 60.64 | 64.53 |
passage_retrieval_en | 70.0 | 84.5 |
passage_retrieval_zh | 56.0 | 70.0 |
trec | 76.5 | 76.5 |
lsht | 43.5 | 45.0 |
Average | 52.48 | 59.20 |