Training#
Most examples we showed in this documentation used pre-trained models as-is. For example, we often used lmsys/vicuna-7b-v1.5 and mistralai/Mistral-7B-Instruct-v0.2 for QA models, and intfloat/e5-base-v2 for our dense embedding models. However, in practice, finetuning these models with task-specific data often leads to much better performance (Karpukhin et al., 2020, Izacard et al., 2021; Wang et al., 2022).
To this end, we provide various training algorithms for both dense retrieval and generative models. This (non-exhaustive and growing) collection of algorithms comes from the latest research in the field, and we provide a simple interface to train these models with your own data.
Algorithm |
Target Model |
Training Data |
Section |
|---|---|---|---|
CTL |
Encoder |
(q, p, hard negative p†) |
|
DCA |
Encoder |
(q, p†) |
|
RPG |
Encoder |
(q, p†, a) |
|
SFT |
Decoder |
(h†, q, p, a) |
|
SwR |
Decoder |
(h†, q, p, a) |
|
FiD |
Encoder-Decoder |
(h†, q, p, a) |
where “h” stands for chat history, “q” stands for a question, “p” stands for a (gold) passage, and “a” stands for a (gold) answer. “†” indicates this element is optional. We implement our trainers based on the Trainer class from the transformers library. This means that most of the arguments are the same as if you are familiar with the transformers library.
Note
Currently, our training scripts assume a single GPU setting, since our evaluation loop assumes all models are loaded on a single GPU. To avoid unexpected behavior with multi-GPU training, we suggest setting --evaluation_strategy=no with our current training scripts.
References:
Vladimir Karpukhin, Barlas O ̆guz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen tau Yih. 2020. Dense passage retrieval for open-domain question answering.
Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. Unsupervised dense information retrieval with contrastive learning.
Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly supervised contrastive pre-training.