The SEConvo data is available for download here. See SEConvo/README.md
for data details.
- Recommended to use
python 3.10
. - Obtain the
trained_si_model
andfaiss_index_v5
(conversation snippet dataset). Placetrained_si_model
in themessage_si_detector/
folder andfaiss_index_v5
in these_attempt_detector/
folder. These models are publicly available but are not detailed here due to the anonymity policy. - Place your Hugging Face token in the
hf_token.txt
file and your OpenAI API key in theopenai_api_key.txt
file. - Install the required packages:
pip install -r requirements.txt
- Run the
ConvoSentinel
script:This script generates:bash run_convosentinel.sh SEConvo/annotated_test.json outputs/message_si_preds.json outputs/message_se_preds.json outputs/conv_se_preds.json
- Message-level SI predictions:
outputs/message_si_preds.json
- Message-level SE predictions:
outputs/message_se_preds.json
- Conversation-level SE predictions:
outputs/conv_se_preds.json
- Message-level SI predictions:
- By default, the script runs with
cuda:0
. Modify this in the script if necessary.