full evaluation not complete

Fin-MPNET-Base (v0.1)

This is a fine-tuned sentence-transformers model: It maps sentences & paragraphs to a 768 dimensional dense vector space and can be used for tasks like clustering or semantic search.

This model aims to be very strong on Financial Document Retrieval Tasks, while trying to maintain as much generalized performance as possible.

	FiQA	SciFact	AmazonReviews	OnlineBankingIntent	ArguAna
fin-mpnet-base	79.91	65.40	29.12	80.25	49.11
all-mpnet-base-v2	49.96	65.57	31.92	81.86	46.52
previous SoTA	56.59	-	-	-	-

v0.1 shows SoTA results on FiQA Test set while other non-financial benchmarks only drop a few small % and improvement in others.

Usage (Sentence-Transformers)

Using this model becomes easy when you have sentence-transformers installed:

pip install -U sentence-transformers

Then you can use the model like this:

from sentence_transformers import SentenceTransformer
sentences = ["This is an example sentence", "Each sentence is converted"]

model = SentenceTransformer('mukaj/fin-mpnet-base')
embeddings = model.encode(sentences)
print(embeddings)

Evaluation Results

Model was evaluated during training only on the new finance QA examples, as such only financial relevant benchmarks were evaluated on for v0.1 [FiQA-2018, BankingClassification77]

The model currently shows the highest FiQA Retrieval score on the test set, on the MTEB Leaderboard (https://huggingface.co/spaces/mteb/leaderboard)

The model will have likely suffered some performance on other benchmarks, i.e. BankingClassification77 has dropped from 81.6 to 80.25, this will be addressed for v0.2 and full evaluation on all sets will be run.

Training

"sentence-transformers/all-mpnet-base-v2" was fine-tuned on 150k+ financial document QA examples using MNR Loss.