![](/luohongyin/SAIL/raw/main/images/sail-7b-logo.png)
Towards Robust Grounded Language Modeling [DEMO] | [WEB]
We answer the following questions:
- Does search engine always improves large language models?
- No. We found that the improvement by applying search engines of LLMs is minimal on several tasks. While search engines retrieve a vast range of up-to-date information, the retrieval results can be disputing or distracting.
- How to improve language models with search engines?
- We fine-tune a large language model (LLaMA-7B) grounded on real search engine outputs. The fine-tuned model can automatically distill the informative search results and flag distracting items. With the search-augmented fine-tuning, our model can be significantly boosted by a search engine, outperforming state-of-the-art chatbots including ChatGPT and Vicuna-13B with much fewer parameters.
Bing fine-tuning LLaMA-7B model with a search-augmented corpus, our SAIL-7b model outperforms ChatGPT and Vicuna-13B on instruction following!
Our method also benefits AI for social good! The SAIL-7B model outperforms LLaMA-7B and Vicuna-13B on hate speec detection, stereotype recognition, and search-grounded fact checking tasks
We construct a search-augmented instruction training set with two steps:
- Collect search results for each instruction and input pair
- Relabel the responses to explicitly specify if a search result is informative of distracting with pretrained entailment models by Luo et al, 2023.
We provide the collect search results, and a complete training corpus can be constructed by simply running
bash data_prep.sh
Note that this process including runing a 350M large language model (RoBERTa or DeBERTa based). This option can be switched in the data_prep.sh
file.
The constructed training set can be used to fine-tuned LLaMA-based models with FastChat. If any tokenization error occurs, try replacing the following files with the code files we provide in this repository:
fastchat/fastchat/train/train.py
<-train.py
fastchat/fastchat/conversations.py
<-conversations.py
The training parameters are provided in train.sh
.
The pretrained SAIL-7B model is based on LLaMA, so the applications of the model and demo should align with LLaMA's GPL-3.0 license.
We build a live demo on the Huggingface Space with Gradio. The demo time outs on 1 min so it cannot process very long texts. With the demo, you could test the instruction following ability of SAIL-7B with or without search augmentation.
We plan to release the Delta weights of the pretrained model before July.
If there is any question, submit an issue or contact hyluo AT mit DOT edu
.