Add application file

app.py

@@ -0,0 +1,50 @@
+import torch
+import gradio as gr
+from transformers import DebertaV2Tokenizer
+from modeling_multitask_bias import MultiTaskBiasModel
+
+REPO = "piyush333/deberta-v3-multitask-bias-detector-mach-1"
+CKPT = "model_dpo_epoch_5.pt"
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+
+label_maps = {
+    "political": {0: "left", 1: "neutral", 2: "right"},
+    "gender": {0: "misogynist", 1: "neutral", 2: "misandrist"},
+    "immigration": {0: "anti", 1: "neutral", 2: "pro"}
+}
+
+def load_model():
+    tok = DebertaV2Tokenizer.from_pretrained("microsoft/deberta-v3-base")
+    mdl = MultiTaskBiasModel()
+    state = torch.hub.load_state_dict_from_url(
+        f"https://huggingface.co/{REPO}/resolve/main/{CKPT}",
+        map_location="cpu"
+    )
+    mdl.load_state_dict(state)
+    mdl.to(DEVICE).eval()
+    return tok, mdl
+
+tokenizer, model = load_model()
+
+def predict(text, task):
+    enc = tokenizer(text, return_tensors="pt", truncation=True, padding=True).to(DEVICE)
+    with torch.no_grad():
+        logits = model(enc["input_ids"], enc["attention_mask"], [task])
+        probs = torch.softmax(logits, dim=1)[0].tolist()
+    classes = [label_maps[task][i] for i in range(3)]
+    pred = classes[int(torch.tensor(probs).argmax())]
+    return pred, {cls: round(float(p), 4) for cls, p in zip(classes, probs)}
+
+demo = gr.Interface(
+    fn=predict,
+    inputs=[
+        gr.Textbox(label="Text", lines=4, placeholder="Enter text..."),
+        gr.Dropdown(["political", "gender", "immigration"], label="Task")
+    ],
+    outputs=[gr.Textbox(label="Prediction"), gr.JSON(label="Probabilities")],
+    title="DeBERTaV3 Multi-Task Bias Detector (Mach-1)",
+    description="Detects stance bias in political, gender, and immigration domains"
+)
+
+if __name__ == "__main__":
+    demo.launch()

modeling_multitask_bias.py

@@ -0,0 +1,25 @@
+import torch
+import torch.nn as nn
+from transformers import DebertaV2Model
+
+class MultiTaskBiasModel(nn.Module):
+    def __init__(self, model_name="microsoft/deberta-v3-base"):
+        super().__init__()
+        self.bert = DebertaV2Model.from_pretrained(model_name)
+        hidden = self.bert.config.hidden_size
+        self.heads = nn.ModuleDict({
+            task: nn.Sequential(
+                nn.Linear(hidden, hidden),
+                nn.ReLU(),
+                nn.Dropout(0.2),
+                nn.Linear(hidden, 3)
+            )
+            for task in ["political", "gender", "immigration"]
+        })
+
+    def forward(self, input_ids, attention_mask, tasks):
+        outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask).last_hidden_state[:, 0]
+        logits = []
+        for i in range(len(tasks)):
+            logits.append(self.heads[tasks[i]](outputs[i].unsqueeze(0)))
+        return torch.cat(logits, dim=0)

requirements.txt

@@ -0,0 +1,3 @@
+torch
+transformers
+gradio