🚀🗣️ deployed beta 🚀🗣️

fallacy-detector

code repository for fallacy-detector application

• fallacy-detector/databricks is resources for databricks workspace and training code.
• fallacy-detector/Fallacy_Detector is a react website deployed via netlify.

training

students at san diego state university constructed individual binary classifiers on 10 different logical fallacies. These were jupyter notebooks using the BertTokenizer from nltk to extract features from labeled data, data is split into train/test sets, then loads a pretrained BERT model with a classification head using AdamW optimizer (preferred for Transformers). Performance metrics are obtained from the test set, the goal is probabilistic multi-classification.

labeled data data is synthetically generated via language learning models -- prompts like "do you know what ad hominem is? can you generate 100 examples in a python list for me? 100 non examples?". We additionally leveraged the coco dataset of logical fallacies for non-synthetic examples.

these notebooks were refactored into a single deployed databricks notebook (see training pipeline) which produces an MLflow model to serve inferences over https. Data was ultimately loaded into hive tables to facilitate simpler ETL (see table initialization)

model serving

Databricks is nicely integrated with MLflow, so we were able to publish necessary artifacts and serve our classifier under the models section. Realistically we'd look into publishing these contianers to a separate registry and leverage an in-house k8s approach since deferring completely to databricks comes with a premium $$$. This is why the current version of this requires a personal access token to eliminate excessive usage.

curl -X POST https://dbc-28e35821-942c.cloud.databricks.com/serving-endpoints/fallacy-classifier/invocations \
  -H "Authorization: Bearer {beta-token}" \
  -H "Content-Type: application/json" \
  -d '{
        "dataframe_split": {
          "columns": ["text"],
          "data": [["You are just a student, do not push your opinions... Opinions are dangerous."]]
        }
      }'

website

The react application exists under fallacy-detector/Fallacy_Detector, for more web-devy specific details see the associated REAME there. The website was deployed on netlify, and the request to databricks is proxied by a netlify function.