Open Language Safety Research

The Open Language Safety Research (OpenLSR) project is a series of research efforts aiming to improve the robustness, inclusion, and alignment of language and conversation AI. The project is led by the Spoken Language Systems (SLS) group at MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and its collaborators. OpenLSR has released the following:

• A Twitter bot that checks the language of the tweet for factualness and fairness using the vicuna-13b large language model (LLM) and a chain-of-thought prompting strategy.
• Sail-7B is a search engine-grounded LLM that generates language grounding on noisy search results, improving both the transparency and robustness of LLMs with distracting information.

OpenLSR has published the following papers accepted at computational linguistics conferences:

• "Logic Against Bias: Textual Entailment Mitigates Stereotypical Sentence Reasoning," Hongyin Luo and James Glass, March 2023.
• "Interpretable Unified Language Checking," Tianhua Zhang, Hongyin Luo, Yung-Sung Chuang, Wei Fang, Luc Gaitskell, Thomas Hartvigsen, Xixin Wu, Danny Fox, Helen Meng, James Glass, April 2023.

The OpenLSR Twitter bot was released on April 13, 2023. Twitter users can reply to a tweet, mentioning @openlsr, to perform a language check that will be automatically posted. The factualness and fairness checking is conducted using the vicuna-13b large language model (LLM) and the chain-of-thought prompting strategy described in the OpenLSR paper "Interpretable Unified Language Checking." The initial release is an experiment, and the language-checking system is imperfect. OpenLSR is collecting data to improve the performance of the current model. The project is advised and supported by professors and colleagues at MIT CSAIL, CUHK CPII, and MIT Linguistics.

Example of the OpenLSR Twitter bot checking a tweet.

On May 24, 2023, OpenLSR released SAIL-7B an LLM fine-tuned on search engine results to generate language grounded on noisy search results. While search engine results return disputed or distracting information, SAIL-7B is trained to automatically identify information results and flag distracting items. With search-augmented fine-tuning, the model's performance can be improved by a search engine. OpenLSR tested the performance of SAIL-7b against other search-augmented models, including ChatGPT and Vicuna, demonstrating better results on open-ended QA benchmarks.

SAIL-7B is based on the LLaMA-7b model with a search-augmented instruction training set. This includes using 52,000 instructions designed by the Alpaca Team with the response generated by GPT-4. In addition, OpenLSR collected ten search results (titles + previews only) for each instruction with DuckDuckGO.com and a BM25-based Wikipedia retriever implemented by Pyserini, feeding the top zero to five sampled search results to LLaMA for fine-tuning and evaluation. The model was trained on 4 NVIDIA RTX A6000 GPUs (4x48GB), taking roughly twenty-four hours.