Hi I work in Legal tech and wanted to see how it did reviewing documents and the results are pretty astounding in terms of cost and speed.

Hi I work in Legal tech and wanted to see how it did reviewing documents and the results are pretty astounding in terms of cost and speed. I ran 9840 documents from the enron dataset through the model with responsive nonresponsive designations and the model performed eye watering...ly fast. 44.6 documents a second 1.05 in cost for 9.8k docs. Was it perfect on the calls? eh it was pretty good given i slammed a prompt in there. With more case specifics and tuning it could be better: Metric Result Accuracy 82.9% Precision 80.6% Recall 80.6% F1 80.6% What was great is that lower confidence coincided with more mistakes (See image) Meaning if we only accepted high confidence(>=95) answers we get these measures: Metric High-confidence answers only Accuracy 93.7% Precision 87.5% Recall within that group 95.5% F1 91.3% Thats incredible with no refinement to the prompt. Exceptional model and cant wait to test