Experimenting with few-shot classification.

app.py

@@ -60,20 +60,20 @@ class OnnxSetFitModel:
 # "Xenova/distilbert-base-uncased-mnli" "typeform/distilbert-base-uncased-mnli" Bad answers
 # "Xenova/deBERTa-v3-base-mnli" "MoritzLaurer/DeBERTa-v3-base-mnli" Still a bit slow and not great answers
 # "xenova/nli-deberta-v3-small" "cross-encoder/nli-deberta-v3-small" Was using this for a good while and it was...okay
-model_name = "Xenova/deBERTa-v3-base-mnli"
-file_name = "onnx/model_quantized.onnx"
-tokenizer_name = "MoritzLaurer/DeBERTa-v3-base-mnli"
+model_name = "MoritzLaurer/bge-m3-zeroshot-v2.0"
+file_name = "onnx/model.onnx"
+tokenizer_name = "MoritzLaurer/bge-m3-zeroshot-v2.0"
 model = ORTModelForSequenceClassification.from_pretrained(model_name, file_name=file_name)
 tokenizer = AutoTokenizer.from_pretrained(tokenizer_name, model_max_length=512)
 classifier = pipeline(task="zero-shot-classification", model=model, tokenizer=tokenizer)
 
-few_shot_tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-small-en-v1.5', model_max_length=512)
-ort_model = ORTModelForFeatureExtraction.from_pretrained('BAAI/bge-small-en-v1.5', file_name="onnx/model.onnx")
-few_shot_model = SetFitModel.from_pretrained("moshew/bge-small-en-v1.5_setfit-sst2-english")
+few_shot_tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-m3', model_max_length=512) # 'BAAI/bge-small-en-v1.5'
+ort_model = ORTModelForFeatureExtraction.from_pretrained('BAAI/bge-m3', file_name="onnx/model.onnx") # 'BAAI/bge-small-en-v1.5'
+few_shot_model = SetFitModel.from_pretrained("BAAI/bge-m3") # "moshew/bge-small-en-v1.5_setfit-sst2-english"
 
 # Train few_shot_model
 candidate_labels = ["correct", "wrong"]
-reference_dataset = load_dataset("emotion")
+reference_dataset = load_dataset("SetFit/sst2")
 dummy_dataset = Dataset.from_dict({})
 train_dataset = get_templated_dataset(dummy_dataset, candidate_labels=candidate_labels, sample_size=8, template="This conclusion is {}.")
 args = TrainingArguments(