Upload app.py

app.py

@@ -0,0 +1,66 @@
+import torch
+import torch.nn as nn
+from transformers import AutoTokenizer, AutoModel
+
+import streamlit as st
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+# ✅ Load tokenizer from current directory
+tokenizer = AutoTokenizer.from_pretrained("./")
+
+# ✅ Define model
+class ScoringModel(nn.Module):
+    def __init__(self, base_model_name="microsoft/deberta-v3-small", dropout_rate=0.242):
+        super().__init__()
+        self.base = AutoModel.from_pretrained(base_model_name)
+        self.base.gradient_checkpointing_enable()
+        self.dropout1 = nn.Dropout(dropout_rate)
+        self.dropout2 = nn.Dropout(dropout_rate)
+        self.dropout3 = nn.Dropout(dropout_rate)
+        self.classifier = nn.Linear(self.base.config.hidden_size, 1)
+
+    def forward(self, input_ids, attention_mask):
+        hidden = self.base(input_ids=input_ids, attention_mask=attention_mask).last_hidden_state[:, 0]
+        logits = (self.classifier(self.dropout1(hidden)) +
+                  self.classifier(self.dropout2(hidden)) +
+                  self.classifier(self.dropout3(hidden))) / 3
+        return logits
+
+# ✅ Load your fine-tuned weights
+model = ScoringModel()
+model.load_state_dict(torch.load("scoring_model.pt", map_location=device))
+model.to(device)
+model.eval()
+
+# ✅ Streamlit UI
+st.title("LLM Response Scoring App 🚀")
+prompt = st.text_area("Enter a prompt:")
+response_a = st.text_area("Response A:")
+response_b = st.text_area("Response B:")
+
+if st.button("Predict Better Response"):
+    text_a = f"Prompt: {prompt} [SEP] {response_a}"
+    text_b = f"Prompt: {prompt} [SEP] {response_b}"
+
+    inputs_a = tokenizer(text_a, return_tensors="pt", padding="max_length", truncation=True, max_length=186)
+    inputs_b = tokenizer(text_b, return_tensors="pt", padding="max_length", truncation=True, max_length=186)
+
+    inputs_a = {k: v.to(device) for k, v in inputs_a.items() if k in ["input_ids", "attention_mask"]}
+    inputs_b = {k: v.to(device) for k, v in inputs_b.items() if k in ["input_ids", "attention_mask"]}
+
+    with torch.no_grad():
+        score_a = model(**inputs_a).squeeze()
+        score_b = model(**inputs_b).squeeze()
+
+    prob_a = torch.sigmoid(score_a).item()
+    prob_b = torch.sigmoid(score_b).item()
+
+    if prob_b > prob_a:
+        st.success("✅ Model predicts: Response B is better")
+    else:
+        st.success("✅ Model predicts: Response A is better")
+
+    st.write(f"🔵 Probability A: {prob_a:.4f}")
+    st.write(f"🟠 Probability B: {prob_b:.4f}")
+