app.py

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer
import gradio as gr
import torch

# Schritt 1: Dataset laden und überprüfen
# Falls "KeyError: 'text'" auftritt, Spaltennamen prüfen

dataset = load_dataset("armanc/scientific_papers", "arxiv", trust_remote_code=True)  # Falls du PubMed nutzt, ersetze "arxiv" mit "pubmed"
print(dataset)

# Schritt 2: Tokenizer vorbereiten
tokenizer = AutoTokenizer.from_pretrained("allenai/scibert_scivocab_uncased")

def tokenize_function(examples):
    return tokenizer(examples["abstract"], padding="max_length", truncation=True, max_length=151)

dataset = dataset.map(tokenize_function, batched=True)

# Schritt 3: Modell laden
model = AutoModelForSequenceClassification.from_pretrained("allenai/scibert_scivocab_uncased", num_labels=3)

# Anpassung für Trainingsdaten: Label-Spalte hinzufügen
def add_labels(example):
    example["labels"] = 1  # Dummy-Label, falls nicht vorhanden (1=positiv, 0=negativ, 2=neutral o.Ä.)
    return example

dataset = dataset.map(add_labels)

# Schritt 4: Trainingsparameter setzen
training_args = TrainingArguments(
    output_dir="./results",
    eval_strategy="epoch",
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8,
    num_train_epochs=3,
    learning_rate=5e-5,
    weight_decay=0.01,
    logging_dir="./logs",
    logging_steps=500,
)

# Schritt 5: Trainer erstellen und Training starten
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=dataset["train"],
    eval_dataset=dataset["validation"]
)
trainer.train()

# Schritt 6: Modell speichern
trainer.save_model("./trained_model")
tokenizer.save_pretrained("./trained_model")

# Schritt 7: Modell für Gradio bereitstellen
def predict(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding="max_length", max_length=151)
    with torch.no_grad():
        outputs = model(**inputs)
    logits = outputs.logits
    probabilities = torch.nn.functional.softmax(logits, dim=-1)
    return {f"Label {i}": float(probabilities[0][i]) for i in range(len(probabilities[0]))}

iface = gr.Interface(
    fn=predict,
    inputs=gr.Textbox(lines=5, placeholder="Paste an abstract here..."),
    outputs=gr.Label(),
    title="Scientific Paper Evaluator",
    description="This AI model scores scientific papers based on relevance, uniqueness, and redundancy."
)

iface.launch()