app.py

# ============================================================================
# ConfliBERT - Conflict & Political Violence NLP Toolkit
# University of Texas at Dallas | Event Data Lab
# ============================================================================

import os
os.environ['TF_ENABLE_ONEDNN_OPTS'] = '0'

# Patch gradio_client bug: bool JSON sub-schemas crash schema parsing
try:
    from gradio_client import utils as _gc_utils
    _original_get_type = _gc_utils.get_type
    def _patched_get_type(schema):
        if isinstance(schema, bool):
            return "Any"
        return _original_get_type(schema)
    _gc_utils.get_type = _patched_get_type

    _original_json_schema = _gc_utils._json_schema_to_python_type
    def _patched_json_schema(schema, defs=None):
        if isinstance(schema, bool):
            return "Any"
        return _original_json_schema(schema, defs)
    _gc_utils._json_schema_to_python_type = _patched_json_schema
except Exception:
    pass

import torch
from transformers import (
    AutoTokenizer,
    AutoModelForSequenceClassification,
    AutoModelForTokenClassification,
    TrainingArguments,
    Trainer,
    EarlyStoppingCallback,
    TrainerCallback,
)

# QA model uses TensorFlow (transformers <5) or PyTorch fallback (transformers >=5)
_USE_TF_QA = False
try:
    import tensorflow as tf
    import tf_keras   # noqa: F401
    import keras      # noqa: F401
    from transformers import TFAutoModelForQuestionAnswering
    _USE_TF_QA = True
except (ImportError, ModuleNotFoundError):
    from transformers import AutoModelForQuestionAnswering
import gradio as gr
import numpy as np
import pandas as pd
import re
import csv
import tempfile
from sklearn.metrics import (
    accuracy_score as sk_accuracy,
    precision_score as sk_precision,
    recall_score as sk_recall,
    f1_score as sk_f1,
    roc_curve,
    auc as sk_auc,
)
from sklearn.preprocessing import label_binarize
from torch.utils.data import Dataset as TorchDataset
import gc

# LoRA / QLoRA support (optional)
try:
    from peft import LoraConfig, get_peft_model, TaskType
    PEFT_AVAILABLE = True
except ImportError:
    PEFT_AVAILABLE = False

try:
    from transformers import BitsAndBytesConfig
    BNB_AVAILABLE = True
except ImportError:
    BNB_AVAILABLE = False


# ============================================================================
# CONFIGURATION
# ============================================================================

if torch.cuda.is_available():
    device = torch.device('cuda')
elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
    device = torch.device('mps')
else:
    device = torch.device('cpu')

MAX_TOKEN_LENGTH = 512


def get_system_info():
    """Build an HTML string describing the user's compute environment."""
    import platform
    lines = []

    # Device
    if device.type == 'cuda':
        gpu_name = torch.cuda.get_device_name(0)
        vram = torch.cuda.get_device_properties(0).total_memory / (1024 ** 3)
        lines.append(f"GPU: {gpu_name} ({vram:.1f} GB VRAM)")
        lines.append("FP16 training: supported")
    elif device.type == 'mps':
        lines.append("GPU: Apple Silicon (MPS)")
        lines.append("FP16 training: not supported on MPS")
    else:
        lines.append("GPU: None detected (using CPU)")
        lines.append("FP16 training: not supported on CPU")

    # CPU / RAM
    import os
    cpu_count = os.cpu_count() or 1
    lines.append(f"CPU cores: {cpu_count}")
    try:
        import psutil
        ram_gb = psutil.virtual_memory().total / (1024 ** 3)
        lines.append(f"RAM: {ram_gb:.1f} GB")
    except ImportError:
        pass

    lines.append(f"Platform: {platform.system()} {platform.machine()}")
    lines.append(f"PyTorch: {torch.__version__}")

    return " · ".join(lines)

FINETUNE_MODELS = {
    "ConfliBERT (recommended for conflict/political text)": "snowood1/ConfliBERT-scr-uncased",
    "BERT Base Uncased": "bert-base-uncased",
    "BERT Base Cased": "bert-base-cased",
    "RoBERTa Base": "roberta-base",
    "ModernBERT Base": "answerdotai/ModernBERT-base",
    "DeBERTa v3 Base": "microsoft/deberta-v3-base",
    "DistilBERT Base Uncased": "distilbert-base-uncased",
}

NER_LABELS = {
    'Organisation': '#3b82f6',
    'Person': '#ef4444',
    'Location': '#10b981',
    'Quantity': '#ff6b35',
    'Weapon': '#8b5cf6',
    'Nationality': '#06b6d4',
    'Temporal': '#ec4899',
    'DocumentReference': '#92400e',
    'MilitaryPlatform': '#f59e0b',
    'Money': '#f472b6',
}

CLASS_NAMES = ['Negative', 'Positive']
MULTI_CLASS_NAMES = ["Armed Assault", "Bombing or Explosion", "Kidnapping", "Other"]


# ============================================================================
# PRETRAINED MODEL LOADING
# ============================================================================

qa_model_name = 'salsarra/ConfliBERT-QA'
if _USE_TF_QA:
    qa_model = TFAutoModelForQuestionAnswering.from_pretrained(qa_model_name)
else:
    qa_model = AutoModelForQuestionAnswering.from_pretrained(qa_model_name, from_tf=True)
qa_tokenizer = AutoTokenizer.from_pretrained(qa_model_name)

ner_model_name = 'eventdata-utd/conflibert-named-entity-recognition'
ner_model = AutoModelForTokenClassification.from_pretrained(ner_model_name).to(device)
ner_tokenizer = AutoTokenizer.from_pretrained(ner_model_name)

clf_model_name = 'eventdata-utd/conflibert-binary-classification'
clf_model = AutoModelForSequenceClassification.from_pretrained(clf_model_name).to(device)
clf_tokenizer = AutoTokenizer.from_pretrained(clf_model_name)

multi_clf_model_name = 'eventdata-utd/conflibert-satp-relevant-multilabel'
multi_clf_model = AutoModelForSequenceClassification.from_pretrained(multi_clf_model_name).to(device)
multi_clf_tokenizer = AutoTokenizer.from_pretrained(multi_clf_model_name)


# ============================================================================
# UTILITY FUNCTIONS
# ============================================================================

def get_path(f):
    """Get file path from Gradio file component output."""
    if f is None:
        return None
    return f if isinstance(f, str) else getattr(f, 'name', str(f))


def truncate_text(text, tokenizer, max_length=MAX_TOKEN_LENGTH):
    tokens = tokenizer.encode(text, truncation=False)
    if len(tokens) > max_length:
        tokens = tokens[:max_length - 1] + [tokenizer.sep_token_id]
        return tokenizer.decode(tokens, skip_special_tokens=True)
    return text


def info_callout(text):
    """Wrap markdown text in a styled callout div to avoid Gradio double-border."""
    return (
        "<div class='info-callout-inner' style='"
        "background: #fff7f3; border-left: 3px solid #ff6b35; "
        "padding: 0.75rem 1rem; border-radius: 0 8px 8px 0; "
        "font-size: 0.9rem;'>\n\n"
        f"{text}\n\n</div>"
    )


def handle_error(e, default_limit=512):
    msg = str(e)
    match = re.search(
        r"The size of tensor a \((\d+)\) must match the size of tensor b \((\d+)\)", msg
    )
    if match:
        return (
            f"<span style='color: #ef4444; font-weight: 600;'>"
            f"Error: Input ({match.group(1)} tokens) exceeds model limit ({match.group(2)})</span>"
        )
    match_qa = re.search(r"indices\[0,(\d+)\] = \d+ is not in \[0, (\d+)\)", msg)
    if match_qa:
        return (
            f"<span style='color: #ef4444; font-weight: 600;'>"
            f"Error: Input too long for model (limit: {match_qa.group(2)} tokens)</span>"
        )
    return f"<span style='color: #ef4444; font-weight: 600;'>Error: {msg}</span>"


# ============================================================================
# INFERENCE FUNCTIONS
# ============================================================================

def question_answering(context, question):
    if not context or not question:
        return "Please provide both context and question."
    try:
        if _USE_TF_QA:
            inputs = qa_tokenizer(question, context, return_tensors='tf', truncation=True)
            outputs = qa_model(inputs)
            start = tf.argmax(outputs.start_logits, axis=1).numpy()[0]
            end = tf.argmax(outputs.end_logits, axis=1).numpy()[0] + 1
            tokens = qa_tokenizer.convert_ids_to_tokens(
                inputs['input_ids'].numpy()[0][start:end]
            )
        else:
            inputs = qa_tokenizer(question, context, return_tensors='pt', truncation=True)
            with torch.no_grad():
                outputs = qa_model(**inputs)
            start = torch.argmax(outputs.start_logits, dim=1).item()
            end = torch.argmax(outputs.end_logits, dim=1).item() + 1
            tokens = qa_tokenizer.convert_ids_to_tokens(
                inputs['input_ids'][0][start:end]
            )
        answer = qa_tokenizer.convert_tokens_to_string(tokens)
        return f"<span style='color: #10b981; font-weight: 600;'>{answer}</span>"
    except Exception as e:
        return handle_error(e)


def named_entity_recognition(text, output_format='html'):
    if not text:
        return "Please provide text for analysis."
    try:
        inputs = ner_tokenizer(text, return_tensors='pt', truncation=True)
        with torch.no_grad():
            outputs = ner_model(**inputs)
        results = outputs.logits.argmax(dim=2).squeeze().tolist()
        tokens = ner_tokenizer.convert_ids_to_tokens(inputs['input_ids'].squeeze().tolist())
        tokens = [t.replace('[UNK]', "'") for t in tokens]

        entities = []
        seen_labels = set()
        current_entity = []
        current_label = None

        for i in range(len(tokens)):
            token = tokens[i]
            label = ner_model.config.id2label[results[i]].split('-')[-1]

            if token.startswith('##'):
                if entities:
                    if output_format == 'html':
                        entities[-1][0] += token[2:]
                    elif current_entity:
                        current_entity[-1] = current_entity[-1] + token[2:]
            else:
                if output_format == 'csv':
                    if label != 'O':
                        if label == current_label:
                            current_entity.append(token)
                        else:
                            if current_entity:
                                entities.append([' '.join(current_entity), current_label])
                            current_entity = [token]
                            current_label = label
                    else:
                        if current_entity:
                            entities.append([' '.join(current_entity), current_label])
                            current_entity = []
                            current_label = None
                else:
                    entities.append([token, label])

                if label != 'O':
                    seen_labels.add(label)

        if output_format == 'csv' and current_entity:
            entities.append([' '.join(current_entity), current_label])

        if output_format == 'csv':
            grouped = {}
            for token, label in entities:
                if label != 'O':
                    grouped.setdefault(label, []).append(token)
            parts = []
            for label, toks in grouped.items():
                unique = list(dict.fromkeys(toks))
                parts.append(f"{label}: {' | '.join(unique)}")
            return ' || '.join(parts)

        # HTML output
        highlighted = ""
        for token, label in entities:
            color = NER_LABELS.get(label, 'inherit')
            if label != 'O':
                highlighted += (
                    f"<span style='color: {color}; font-weight: 600;'>{token}</span> "
                )
            else:
                highlighted += f"{token} "

        if seen_labels:
            legend_items = ""
            for label in sorted(seen_labels):
                color = NER_LABELS.get(label, '#666')
                legend_items += (
                    f"<li style='color: {color}; font-weight: 600; "
                    f"background: {color}15; padding: 2px 8px; border-radius: 4px; "
                    f"font-size: 0.85rem;'>{label}</li>"
                )
            legend = (
                f"<div style='margin-top: 1rem; padding-top: 0.75rem; "
                f"border-top: 1px solid #e5e7eb;'>"
                f"<strong>Entities found:</strong>"
                f"<ul style='list-style: none; padding: 0; display: flex; "
                f"flex-wrap: wrap; gap: 0.5rem; margin-top: 0.5rem;'>"
                f"{legend_items}</ul></div>"
            )
            return f"<div style='line-height: 1.8;'>{highlighted}</div>{legend}"
        else:
            return (
                f"<div style='line-height: 1.8;'>{highlighted}</div>"
                f"<div style='color: #888; margin-top: 0.5rem;'>No entities detected.</div>"
            )

    except Exception as e:
        return handle_error(e)


def predict_with_model(text, model, tokenizer):
    """Run inference with an arbitrary classification model."""
    model.eval()
    dev = next(model.parameters()).device
    inputs = tokenizer(
        text, return_tensors='pt', truncation=True, padding=True, max_length=512
    )
    inputs = {k: v.to(dev) for k, v in inputs.items()}

    with torch.no_grad():
        outputs = model(**inputs)

    probs = torch.softmax(outputs.logits, dim=1).squeeze()
    predicted = torch.argmax(probs).item()
    num_classes = probs.shape[0] if probs.dim() > 0 else 1

    lines = []
    for i in range(num_classes):
        p = probs[i].item() * 100 if probs.dim() > 0 else probs.item() * 100
        if i == predicted:
            lines.append(
                f"<span style='color: #10b981; font-weight: 600;'>"
                f"Class {i}: {p:.2f}% (predicted)</span>"
            )
        else:
            lines.append(f"<span style='color: #9ca3af;'>Class {i}: {p:.2f}%</span>")
    return "<br>".join(lines)


def text_classification(text, custom_model=None, custom_tokenizer=None):
    if not text:
        return "Please provide text for classification."
    try:
        # Use custom model if loaded
        if custom_model is not None and custom_tokenizer is not None:
            return predict_with_model(text, custom_model, custom_tokenizer)

        # Pretrained binary classifier
        inputs = clf_tokenizer(
            text, return_tensors='pt', truncation=True, padding=True
        ).to(device)
        with torch.no_grad():
            outputs = clf_model(**inputs)
        predicted = torch.argmax(outputs.logits, dim=1).item()
        confidence = torch.softmax(outputs.logits, dim=1).max().item() * 100

        if predicted == 1:
            return (
                f"<span style='color: #10b981; font-weight: 600;'>"
                f"Positive -- Related to conflict, violence, or politics. "
                f"(Confidence: {confidence:.1f}%)</span>"
            )
        else:
            return (
                f"<span style='color: #ef4444; font-weight: 600;'>"
                f"Negative -- Not related to conflict, violence, or politics. "
                f"(Confidence: {confidence:.1f}%)</span>"
            )
    except Exception as e:
        return handle_error(e)


def multilabel_classification(text, custom_model=None, custom_tokenizer=None):
    if not text:
        return "Please provide text for classification."
    try:
        # Use custom model if loaded
        if custom_model is not None and custom_tokenizer is not None:
            return predict_with_model(text, custom_model, custom_tokenizer)

        inputs = multi_clf_tokenizer(
            text, return_tensors='pt', truncation=True, padding=True
        ).to(device)
        with torch.no_grad():
            outputs = multi_clf_model(**inputs)
        probs = torch.sigmoid(outputs.logits).squeeze().tolist()

        results = []
        for i in range(len(probs)):
            conf = probs[i] * 100
            if probs[i] >= 0.5:
                results.append(
                    f"<span style='color: #10b981; font-weight: 600;'>"
                    f"{MULTI_CLASS_NAMES[i]}: {conf:.1f}%</span>"
                )
            else:
                results.append(
                    f"<span style='color: #9ca3af;'>"
                    f"{MULTI_CLASS_NAMES[i]}: {conf:.1f}%</span>"
                )
        return "<br>".join(results)
    except Exception as e:
        return handle_error(e)


# ============================================================================
# CSV BATCH PROCESSING
# ============================================================================

def process_csv_ner(file):
    path = get_path(file)
    if path is None:
        return None
    df = pd.read_csv(path)
    if 'text' not in df.columns:
        raise ValueError("CSV must contain a 'text' column")

    entities = []
    for text in df['text']:
        if pd.isna(text):
            entities.append("")
        else:
            entities.append(named_entity_recognition(str(text), output_format='csv'))
    df['entities'] = entities

    out = tempfile.NamedTemporaryFile(suffix='_ner_results.csv', delete=False)
    df.to_csv(out.name, index=False)
    return out.name


def process_csv_binary(file, custom_model=None, custom_tokenizer=None):
    path = get_path(file)
    if path is None:
        return None
    df = pd.read_csv(path)
    if 'text' not in df.columns:
        raise ValueError("CSV must contain a 'text' column")

    results = []
    for text in df['text']:
        if pd.isna(text):
            results.append("")
        else:
            html = text_classification(str(text), custom_model, custom_tokenizer)
            results.append(re.sub(r'<[^>]+>', '', html).strip())
    df['classification_results'] = results

    out = tempfile.NamedTemporaryFile(suffix='_classification_results.csv', delete=False)
    df.to_csv(out.name, index=False)
    return out.name


def process_csv_multilabel(file):
    path = get_path(file)
    if path is None:
        return None
    df = pd.read_csv(path)
    if 'text' not in df.columns:
        raise ValueError("CSV must contain a 'text' column")

    results = []
    for text in df['text']:
        if pd.isna(text):
            results.append("")
        else:
            html = multilabel_classification(str(text))
            results.append(re.sub(r'<[^>]+>', '', html).strip())
    df['multilabel_results'] = results

    out = tempfile.NamedTemporaryFile(suffix='_multilabel_results.csv', delete=False)
    df.to_csv(out.name, index=False)
    return out.name


def process_csv_qa(file):
    path = get_path(file)
    if path is None:
        return None
    df = pd.read_csv(path)
    if 'context' not in df.columns or 'question' not in df.columns:
        raise ValueError("CSV must contain 'context' and 'question' columns")

    answers = []
    for _, row in df.iterrows():
        if pd.isna(row['context']) or pd.isna(row['question']):
            answers.append("")
        else:
            html = question_answering(str(row['context']), str(row['question']))
            answers.append(re.sub(r'<[^>]+>', '', html).strip())
    df['answer'] = answers

    out = tempfile.NamedTemporaryFile(suffix='_qa_results.csv', delete=False)
    df.to_csv(out.name, index=False)
    return out.name


# ============================================================================
# FINETUNING
# ============================================================================

class TextClassificationDataset(TorchDataset):
    """PyTorch Dataset for text classification with HuggingFace tokenizers."""

    def __init__(self, texts, labels, tokenizer, max_length=512):
        self.encodings = tokenizer(
            texts, truncation=True, padding=True,
            max_length=max_length, return_tensors=None,
        )
        self.labels = labels

    def __getitem__(self, idx):
        item = {k: torch.tensor(v[idx]) for k, v in self.encodings.items()}
        item['labels'] = torch.tensor(self.labels[idx], dtype=torch.long)
        return item

    def __len__(self):
        return len(self.labels)


def parse_data_file(file_path):
    """Parse a TSV/CSV data file. Expected format: text<separator>label (no header).
    Labels must be integers. Returns (texts, labels, num_labels)."""
    path = get_path(file_path)
    texts, labels = [], []

    # Detect delimiter from first line
    with open(path, 'r', encoding='utf-8') as f:
        first_line = f.readline()
    delimiter = '\t' if '\t' in first_line else ','

    with open(path, 'r', encoding='utf-8') as f:
        reader = csv.reader(f, delimiter=delimiter, quotechar='"')
        for row in reader:
            if len(row) < 2:
                continue
            try:
                label = int(row[-1].strip())
                text = row[0].strip() if len(row) == 2 else delimiter.join(row[:-1]).strip()
                if text:
                    texts.append(text)
                    labels.append(label)
            except (ValueError, IndexError):
                continue  # skip header or malformed rows

    if not texts:
        raise ValueError(
            "No valid data rows found. Expected format: text<tab>label (no header row)"
        )

    num_labels = max(labels) + 1
    return texts, labels, num_labels


class LogCallback(TrainerCallback):
    """Captures training logs for display in the UI."""

    def __init__(self):
        self.entries = []

    def on_log(self, args, state, control, logs=None, **kwargs):
        if logs:
            self.entries.append({**logs})

    def format(self):
        lines = []
        skip_keys = {
            'total_flos', 'train_runtime', 'train_samples_per_second',
            'train_steps_per_second', 'train_loss',
        }
        for entry in self.entries:
            parts = []
            for k, v in sorted(entry.items()):
                if k in skip_keys:
                    continue
                if isinstance(v, float):
                    parts.append(f"{k}: {v:.4f}")
                elif isinstance(v, (int, np.integer)):
                    parts.append(f"{k}: {v}")
            if parts:
                lines.append("  ".join(parts))
        return "\n".join(lines)


def make_compute_metrics(task_type):
    """Factory for compute_metrics function based on task type."""

    def compute_metrics(eval_pred):
        logits, labels = eval_pred
        preds = np.argmax(logits, axis=-1)
        acc = sk_accuracy(labels, preds)

        if task_type == "Binary":
            return {
                'accuracy': acc,
                'precision': sk_precision(labels, preds, zero_division=0),
                'recall': sk_recall(labels, preds, zero_division=0),
                'f1': sk_f1(labels, preds, zero_division=0),
            }
        else:
            return {
                'accuracy': acc,
                'f1_macro': sk_f1(labels, preds, average='macro', zero_division=0),
                'f1_micro': sk_f1(labels, preds, average='micro', zero_division=0),
                'precision_macro': sk_precision(
                    labels, preds, average='macro', zero_division=0
                ),
                'precision_micro': sk_precision(
                    labels, preds, average='micro', zero_division=0
                ),
                'recall_macro': sk_recall(
                    labels, preds, average='macro', zero_division=0
                ),
                'recall_micro': sk_recall(
                    labels, preds, average='micro', zero_division=0
                ),
            }

    return compute_metrics


def run_finetuning(
    train_file, dev_file, test_file, task_type, model_display_name,
    epochs, batch_size, lr, weight_decay, warmup_ratio, max_seq_len,
    grad_accum, fp16, patience, scheduler,
    use_lora, lora_rank, lora_alpha, use_qlora,
    progress=gr.Progress(track_tqdm=True),
):
    """Main finetuning function. Returns logs, metrics, model state, and visibility updates."""
    try:
        # Validate inputs
        if train_file is None or dev_file is None or test_file is None:
            raise ValueError("Please upload all three data files (train, dev, test).")

        epochs = int(epochs)
        batch_size = int(batch_size)
        max_seq_len = int(max_seq_len)
        grad_accum = int(grad_accum)
        patience = int(patience)

        # Parse data files
        train_texts, train_labels, n_train = parse_data_file(train_file)
        dev_texts, dev_labels, n_dev = parse_data_file(dev_file)
        test_texts, test_labels, n_test = parse_data_file(test_file)

        num_labels = max(n_train, n_dev, n_test)
        if task_type == "Binary" and num_labels > 2:
            raise ValueError(
                f"Binary task selected but found {num_labels} label classes in data. "
                f"Use Multiclass instead."
            )
        if task_type == "Binary":
            num_labels = 2

        # Load model and tokenizer
        model_id = FINETUNE_MODELS[model_display_name]
        tokenizer = AutoTokenizer.from_pretrained(model_id)

        lora_active = False
        if use_qlora:
            if not (PEFT_AVAILABLE and BNB_AVAILABLE and torch.cuda.is_available()):
                raise ValueError(
                    "QLoRA requires a CUDA GPU and the peft + bitsandbytes packages."
                )
            bnb_config = BitsAndBytesConfig(
                load_in_4bit=True,
                bnb_4bit_quant_type="nf4",
                bnb_4bit_compute_dtype=torch.float16,
                bnb_4bit_use_double_quant=True,
            )
            model = AutoModelForSequenceClassification.from_pretrained(
                model_id, num_labels=num_labels, quantization_config=bnb_config,
            )
        else:
            model = AutoModelForSequenceClassification.from_pretrained(
                model_id, num_labels=num_labels,
            )

        if use_lora or use_qlora:
            if not PEFT_AVAILABLE:
                raise ValueError(
                    "LoRA requires the 'peft' package. Install: pip install peft"
                )
            lora_config = LoraConfig(
                task_type=TaskType.SEQ_CLS,
                r=int(lora_rank),
                lora_alpha=int(lora_alpha),
                lora_dropout=0.1,
                bias="none",
            )
            model.enable_input_require_grads()
            model = get_peft_model(model, lora_config)
            lora_active = True

        # Create datasets
        train_ds = TextClassificationDataset(
            train_texts, train_labels, tokenizer, max_seq_len
        )
        dev_ds = TextClassificationDataset(
            dev_texts, dev_labels, tokenizer, max_seq_len
        )
        test_ds = TextClassificationDataset(
            test_texts, test_labels, tokenizer, max_seq_len
        )

        # Output directory
        output_dir = tempfile.mkdtemp(prefix='conflibert_ft_')

        # Training arguments
        best_metric = 'f1' if task_type == 'Binary' else 'f1_macro'
        training_args = TrainingArguments(
            output_dir=output_dir,
            num_train_epochs=epochs,
            per_device_train_batch_size=batch_size,
            per_device_eval_batch_size=batch_size * 2,
            learning_rate=lr,
            weight_decay=weight_decay,
            warmup_ratio=warmup_ratio,
            gradient_accumulation_steps=grad_accum,
            fp16=fp16 and torch.cuda.is_available(),
            eval_strategy='epoch',
            save_strategy='epoch',
            load_best_model_at_end=True,
            metric_for_best_model=best_metric,
            greater_is_better=True,
            logging_steps=10,
            save_total_limit=2,
            lr_scheduler_type=scheduler,
            report_to='none',
            seed=42,
        )

        # Callbacks
        log_callback = LogCallback()
        callbacks = [log_callback]
        if patience > 0:
            callbacks.append(EarlyStoppingCallback(early_stopping_patience=patience))

        # Create Trainer
        trainer = Trainer(
            model=model,
            args=training_args,
            train_dataset=train_ds,
            eval_dataset=dev_ds,
            compute_metrics=make_compute_metrics(task_type),
            callbacks=callbacks,
        )

        # Train
        train_result = trainer.train()

        # Evaluate on test set
        test_results = trainer.evaluate(test_ds, metric_key_prefix='test')

        # Build log text
        lora_info = ""
        if lora_active:
            method = "QLoRA (4-bit)" if use_qlora else "LoRA"
            lora_info = f"PEFT:  {method}  r={int(lora_rank)}  alpha={int(lora_alpha)}\n"
        header = (
            f"=== Configuration ===\n"
            f"Model: {model_display_name}\n"
            f"       {model_id}\n"
            f"Task:  {task_type} Classification ({num_labels} classes)\n"
            f"Data:  {len(train_texts)} train / {len(dev_texts)} dev / {len(test_texts)} test\n"
            f"Epochs: {epochs}  Batch: {batch_size}  LR: {lr}  Scheduler: {scheduler}\n"
            f"{lora_info}"
            f"\n=== Training Log ===\n"
        )
        runtime = train_result.metrics.get('train_runtime', 0)
        footer = (
            f"\n=== Training Complete ===\n"
            f"Time: {runtime:.1f}s ({runtime / 60:.1f} min)\n"
        )
        log_text = header + log_callback.format() + footer

        # Build metrics DataFrame
        metrics_data = []
        for k, v in sorted(test_results.items()):
            if isinstance(v, (int, float, np.floating, np.integer)) and k != 'test_epoch':
                name = k.replace('test_', '').replace('_', ' ').title()
                metrics_data.append([name, f"{float(v):.4f}"])
        metrics_df = pd.DataFrame(metrics_data, columns=['Metric', 'Score'])

        # Merge LoRA weights back into base model for clean save/inference
        trained_model = trainer.model
        if lora_active and hasattr(trained_model, 'merge_and_unload'):
            trained_model = trained_model.merge_and_unload()
        trained_model = trained_model.cpu()
        trained_model.eval()

        return (
            log_text, metrics_df, trained_model, tokenizer, num_labels,
            gr.Column(visible=True), gr.Column(visible=True),
        )

    except Exception as e:
        error_log = f"Training failed:\n{str(e)}"
        empty_df = pd.DataFrame(columns=['Metric', 'Score'])
        return (
            error_log, empty_df, None, None, None,
            gr.Column(visible=False), gr.Column(visible=False),
        )


# ============================================================================
# MODEL MANAGEMENT (predict, save, load)
# ============================================================================

def predict_finetuned(text, model_state, tokenizer_state, num_labels_state):
    """Run prediction with the finetuned model stored in gr.State."""
    if not text:
        return "Please enter some text."
    if model_state is None:
        return "No model available. Please train a model first."
    return predict_with_model(text, model_state, tokenizer_state)


def save_finetuned_model(model_state, tokenizer_state):
    """Save the finetuned model as a downloadable zip file."""
    if model_state is None:
        return None, "No model to save. Please train a model first."
    try:
        save_dir = tempfile.mkdtemp(prefix='conflibert_save_')
        model_state.save_pretrained(save_dir)
        tokenizer_state.save_pretrained(save_dir)
        import shutil
        zip_path = os.path.join(tempfile.gettempdir(), 'finetuned_model')
        shutil.make_archive(zip_path, 'zip', save_dir)
        return zip_path + '.zip', "Model ready for download."
    except Exception as e:
        return None, f"Error saving model: {str(e)}"


def load_custom_model(path):
    """Load a finetuned classification model from disk."""
    if not path or not os.path.isdir(path):
        return None, None, "Invalid path. Please enter a valid model directory."
    try:
        tokenizer = AutoTokenizer.from_pretrained(path)
        model = AutoModelForSequenceClassification.from_pretrained(path)
        model.eval()
        n = model.config.num_labels
        return model, tokenizer, f"Loaded model with {n} classes from: {path}"
    except Exception as e:
        return None, None, f"Error loading model: {str(e)}"


def reset_custom_model():
    """Reset to the pretrained ConfliBERT binary classifier."""
    return None, None, "Reset to pretrained ConfliBERT binary classifier."


def batch_predict_finetuned(file, model_state, tokenizer_state, num_labels_state):
    """Run batch predictions on a CSV using the finetuned model."""
    if model_state is None:
        return None
    path = get_path(file)
    if path is None:
        return None

    df = pd.read_csv(path)
    if 'text' not in df.columns:
        raise ValueError("CSV must contain a 'text' column")

    model_state.eval()
    dev = next(model_state.parameters()).device

    predictions, confidences = [], []
    for text in df['text']:
        if pd.isna(text):
            predictions.append("")
            confidences.append("")
            continue

        inputs = tokenizer_state(
            str(text), return_tensors='pt', truncation=True,
            padding=True, max_length=512,
        )
        inputs = {k: v.to(dev) for k, v in inputs.items()}
        with torch.no_grad():
            outputs = model_state(**inputs)
        probs = torch.softmax(outputs.logits, dim=1).squeeze()
        pred = torch.argmax(probs).item()
        conf = probs[pred].item() * 100
        predictions.append(str(pred))
        confidences.append(f"{conf:.1f}%")

    df['predicted_class'] = predictions
    df['confidence'] = confidences

    out = tempfile.NamedTemporaryFile(suffix='_predictions.csv', delete=False)
    df.to_csv(out.name, index=False)
    return out.name


EXAMPLES_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "examples")


def load_example_binary():
    """Load the binary classification example dataset."""
    return (
        os.path.join(EXAMPLES_DIR, "binary", "train.tsv"),
        os.path.join(EXAMPLES_DIR, "binary", "dev.tsv"),
        os.path.join(EXAMPLES_DIR, "binary", "test.tsv"),
        "Binary",
    )


def load_example_multiclass():
    """Load the multiclass classification example dataset."""
    return (
        os.path.join(EXAMPLES_DIR, "multiclass", "train.tsv"),
        os.path.join(EXAMPLES_DIR, "multiclass", "dev.tsv"),
        os.path.join(EXAMPLES_DIR, "multiclass", "test.tsv"),
        "Multiclass",
    )


# ============================================================================
# ACTIVE LEARNING
# ============================================================================

def parse_pool_file(file_path):
    """Parse an unlabeled text pool. Accepts CSV with 'text' column, or one text per line."""
    path = get_path(file_path)
    # Try CSV/TSV with 'text' column first
    try:
        df = pd.read_csv(path)
        if 'text' in df.columns:
            texts = [str(t) for t in df['text'].dropna().tolist()]
            if texts:
                return texts
    except Exception:
        pass
    # Fallback: one text per line
    texts = []
    with open(path, 'r', encoding='utf-8') as f:
        for line in f:
            line = line.strip()
            if line:
                texts.append(line)
    if not texts:
        raise ValueError("No texts found in pool file.")
    return texts


def compute_uncertainty(model, tokenizer, texts, strategy='entropy',
                        max_seq_len=512, batch_size=32):
    """Compute uncertainty scores for unlabeled texts. Higher = more uncertain."""
    model.eval()
    dev = next(model.parameters()).device
    scores = []

    for i in range(0, len(texts), batch_size):
        batch_texts = texts[i:i + batch_size]
        inputs = tokenizer(
            batch_texts, return_tensors='pt', truncation=True,
            padding=True, max_length=max_seq_len,
        )
        inputs = {k: v.to(dev) for k, v in inputs.items()}
        with torch.no_grad():
            logits = model(**inputs).logits
        probs = torch.softmax(logits, dim=1).cpu().numpy()

        if strategy == 'entropy':
            s = -np.sum(probs * np.log(probs + 1e-10), axis=1)
        elif strategy == 'margin':
            sorted_p = np.sort(probs, axis=1)
            s = -(sorted_p[:, -1] - sorted_p[:, -2])
        else:  # least_confidence
            s = -np.max(probs, axis=1)
        scores.extend(s.tolist())

    return scores


def _build_al_metrics_chart(metrics_history, task_type):
    """Build a Plotly chart of active-learning metrics across rounds."""
    import plotly.graph_objects as go

    if not metrics_history:
        return None

    rounds = [m['round'] for m in metrics_history]
    train_sizes = [m.get('train_size', 0) for m in metrics_history]

    metric_keys = (['f1', 'accuracy', 'precision', 'recall']
                    if task_type == 'Binary'
                    else ['f1_macro', 'accuracy'])

    fig = go.Figure()
    colors = ['#ff6b35', '#3b82f6', '#10b981', '#8b5cf6']

    for i, key in enumerate(metric_keys):
        values = [m.get(key) for m in metrics_history]
        if any(v is not None for v in values):
            fig.add_trace(go.Scatter(
                x=rounds, y=values, mode='lines+markers',
                name=key.replace('_', ' ').title(),
                line=dict(color=colors[i % len(colors)], width=2),
            ))

    fig.add_trace(go.Bar(
        x=rounds, y=train_sizes, name='Train Size',
        marker_color='rgba(200,200,200,0.4)', yaxis='y2',
    ))

    fig.update_layout(
        xaxis_title='Round', yaxis_title='Score', yaxis_range=[0, 1.05],
        yaxis2=dict(title='Train Size', overlaying='y', side='right'),
        template='plotly_white',
        legend=dict(orientation='h', yanchor='bottom', y=1.02, xanchor='right', x=1),
        height=350, margin=dict(t=40, b=40),
    )
    return fig


def _train_al_model(texts, labels, num_labels, dev_texts, dev_labels,
                    task_type, model_id, epochs, batch_size, lr, max_seq_len,
                    use_lora, lora_rank, lora_alpha):
    """Train a model for one active-learning round. Returns (model, tokenizer, eval_metrics)."""
    tokenizer = AutoTokenizer.from_pretrained(model_id)
    model = AutoModelForSequenceClassification.from_pretrained(
        model_id, num_labels=num_labels,
    )

    if use_lora and PEFT_AVAILABLE:
        lora_cfg = LoraConfig(
            task_type=TaskType.SEQ_CLS,
            r=int(lora_rank), lora_alpha=int(lora_alpha),
            lora_dropout=0.1, bias="none",
        )
        model.enable_input_require_grads()
        model = get_peft_model(model, lora_cfg)

    train_ds = TextClassificationDataset(texts, labels, tokenizer, max_seq_len)
    dev_ds = None
    if dev_texts is not None:
        dev_ds = TextClassificationDataset(dev_texts, dev_labels, tokenizer, max_seq_len)

    output_dir = tempfile.mkdtemp(prefix='conflibert_al_')
    training_args = TrainingArguments(
        output_dir=output_dir,
        num_train_epochs=epochs,
        per_device_train_batch_size=batch_size,
        per_device_eval_batch_size=batch_size * 2,
        learning_rate=lr,
        weight_decay=0.01,
        warmup_ratio=0.1,
        eval_strategy='epoch' if dev_ds else 'no',
        save_strategy='no',
        logging_steps=10,
        report_to='none',
        seed=42,
    )

    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=train_ds,
        eval_dataset=dev_ds,
        compute_metrics=make_compute_metrics(task_type) if dev_ds else None,
    )
    trainer.train()

    eval_metrics = {}
    if dev_ds:
        results = trainer.evaluate()
        for k, v in results.items():
            if isinstance(v, (int, float, np.floating)):
                eval_metrics[k.replace('eval_', '')] = round(float(v), 4)

    trained_model = trainer.model
    if use_lora and PEFT_AVAILABLE and hasattr(trained_model, 'merge_and_unload'):
        trained_model = trained_model.merge_and_unload()

    return trained_model, tokenizer, eval_metrics


def al_initialize(
    seed_file, pool_file, dev_file, task_type, model_display_name,
    query_strategy, query_size, epochs, batch_size, lr, max_seq_len,
    use_lora, lora_rank, lora_alpha,
    progress=gr.Progress(track_tqdm=True),
):
    """Initialize active learning: train on seed data, query first uncertain batch."""
    try:
        if seed_file is None or pool_file is None:
            raise ValueError("Upload both a labeled seed file and an unlabeled pool file.")

        seed_texts, seed_labels, num_labels = parse_data_file(seed_file)
        pool_texts = parse_pool_file(pool_file)

        dev_texts, dev_labels = None, None
        if dev_file is not None:
            dev_texts, dev_labels, _ = parse_data_file(dev_file)

        if task_type == "Binary":
            num_labels = 2

        query_size = int(query_size)
        model_id = FINETUNE_MODELS[model_display_name]

        trained_model, tokenizer, eval_metrics = _train_al_model(
            seed_texts, seed_labels, num_labels, dev_texts, dev_labels,
            task_type, model_id, int(epochs), int(batch_size), lr,
            int(max_seq_len), use_lora, lora_rank, lora_alpha,
        )

        # Build round-0 metrics
        round_metrics = {'round': 0, 'train_size': len(seed_texts)}
        round_metrics.update(eval_metrics)

        # Query uncertain samples from pool
        scores = compute_uncertainty(
            trained_model, tokenizer, pool_texts, query_strategy, int(max_seq_len),
        )
        top_indices = np.argsort(scores)[-query_size:][::-1].tolist()
        query_texts_batch = [pool_texts[i] for i in top_indices]

        annotation_df = pd.DataFrame({
            'Text': query_texts_batch,
            'Label': [''] * len(query_texts_batch),
        })

        al_state = {
            'labeled_texts': list(seed_texts),
            'labeled_labels': list(seed_labels),
            'pool_texts': pool_texts,
            'pool_available': [i for i in range(len(pool_texts)) if i not in set(top_indices)],
            'current_query_indices': top_indices,
            'dev_texts': dev_texts,
            'dev_labels': dev_labels,
            'num_labels': num_labels,
            'round': 1,
            'metrics_history': [round_metrics],
            'model_id': model_id,
            'model_display_name': model_display_name,
            'task_type': task_type,
            'query_strategy': query_strategy,
            'query_size': query_size,
            'epochs': int(epochs),
            'batch_size': int(batch_size),
            'lr': lr,
            'max_seq_len': int(max_seq_len),
            'use_lora': use_lora,
            'lora_rank': int(lora_rank) if use_lora else 8,
            'lora_alpha': int(lora_alpha) if use_lora else 16,
        }

        trained_model = trained_model.cpu()
        trained_model.eval()

        log_text = (
            f"=== Active Learning Initialized ===\n"
            f"Seed: {len(seed_texts)} labeled  |  Pool: {len(pool_texts)} unlabeled\n"
            f"Model: {model_display_name}\n"
            f"Strategy: {query_strategy}  |  Samples/round: {query_size}\n\n"
            f"--- Round 0 (seed) ---\n"
            f"Train size: {len(seed_texts)}\n"
        )
        for k, v in eval_metrics.items():
            log_text += f"  {k}: {v}\n"
        log_text += (
            f"\n--- Round 1: {len(query_texts_batch)} samples queried ---\n"
            f"Label the samples below, then click 'Submit Labels & Next Round'.\n"
        )

        chart = _build_al_metrics_chart([round_metrics], task_type)

        return (
            al_state, trained_model, tokenizer,
            annotation_df, log_text, chart,
            gr.Column(visible=True),
        )

    except Exception as e:
        return (
            {}, None, None,
            pd.DataFrame(columns=['Text', 'Label']),
            f"Initialization failed:\n{str(e)}",
            None,
            gr.Column(visible=False),
        )


def al_submit_and_continue(
    annotation_df, al_state, al_model, al_tokenizer, prev_log,
    progress=gr.Progress(track_tqdm=True),
):
    """Accept user labels, retrain, query next uncertain batch."""
    try:
        if not al_state or al_model is None:
            raise ValueError("No active session. Initialize first.")

        new_texts = annotation_df['Text'].tolist()
        new_labels = []
        for i, raw in enumerate(annotation_df['Label'].tolist()):
            s = str(raw).strip()
            if s in ('', 'nan'):
                raise ValueError(f"Row {i + 1} has no label. Label all samples first.")
            new_labels.append(int(s))

        num_labels = al_state['num_labels']
        for l in new_labels:
            if l < 0 or l >= num_labels:
                raise ValueError(f"Label {l} out of range [0, {num_labels - 1}].")

        # Add newly labeled samples
        al_state['labeled_texts'].extend(new_texts)
        al_state['labeled_labels'].extend(new_labels)

        queried_set = set(al_state['current_query_indices'])
        al_state['pool_available'] = [
            i for i in al_state['pool_available'] if i not in queried_set
        ]

        current_round = al_state['round']

        # Retrain on all labeled data
        trained_model, tokenizer, eval_metrics = _train_al_model(
            al_state['labeled_texts'], al_state['labeled_labels'],
            num_labels, al_state['dev_texts'], al_state['dev_labels'],
            al_state['task_type'], al_state['model_id'],
            al_state['epochs'], al_state['batch_size'], al_state['lr'],
            al_state['max_seq_len'], al_state['use_lora'],
            al_state['lora_rank'], al_state['lora_alpha'],
        )

        round_metrics = {
            'round': current_round,
            'train_size': len(al_state['labeled_texts']),
        }
        round_metrics.update(eval_metrics)
        al_state['metrics_history'].append(round_metrics)

        # Query next batch from remaining pool
        remaining_pool = al_state['pool_available']
        remaining_texts = [al_state['pool_texts'][i] for i in remaining_pool]

        log_add = (
            f"\n--- Round {current_round} complete ---\n"
            f"Added {len(new_labels)} labels  |  "
            f"Total train: {len(al_state['labeled_texts'])}\n"
        )
        for k, v in eval_metrics.items():
            log_add += f"  {k}: {v}\n"

        if remaining_texts:
            scores = compute_uncertainty(
                trained_model, tokenizer, remaining_texts,
                al_state['query_strategy'], al_state['max_seq_len'],
            )
            q = min(al_state['query_size'], len(remaining_texts))
            top_local = np.argsort(scores)[-q:][::-1].tolist()
            top_pool_indices = [remaining_pool[i] for i in top_local]
            query_texts = [al_state['pool_texts'][i] for i in top_pool_indices]

            al_state['current_query_indices'] = top_pool_indices
            al_state['round'] = current_round + 1

            annotation_out = pd.DataFrame({
                'Text': query_texts,
                'Label': [''] * len(query_texts),
            })
            pool_left = len(remaining_pool) - len(top_pool_indices)
            log_add += (
                f"Pool remaining: {pool_left}\n"
                f"\n--- Round {current_round + 1}: {len(query_texts)} samples queried ---\n"
            )
        else:
            annotation_out = pd.DataFrame(columns=['Text', 'Label'])
            al_state['current_query_indices'] = []
            al_state['round'] = current_round + 1
            log_add += "\nPool exhausted. Active learning complete!\n"

        trained_model = trained_model.cpu()
        trained_model.eval()

        chart = _build_al_metrics_chart(al_state['metrics_history'], al_state['task_type'])
        log_text = prev_log + log_add

        return (
            al_state, trained_model, tokenizer,
            annotation_out, log_text, chart,
        )

    except Exception as e:
        return (
            al_state, al_model, al_tokenizer,
            pd.DataFrame(columns=['Text', 'Label']),
            prev_log + f"\nError: {str(e)}\n",
            None,
        )


def al_save_model(al_model, al_tokenizer):
    """Save the active-learning model as a downloadable zip file."""
    if al_model is None:
        return None, "No model to save. Run at least one round first."
    try:
        save_dir = tempfile.mkdtemp(prefix='conflibert_al_save_')
        al_model.save_pretrained(save_dir)
        al_tokenizer.save_pretrained(save_dir)
        import shutil
        zip_path = os.path.join(tempfile.gettempdir(), 'al_model')
        shutil.make_archive(zip_path, 'zip', save_dir)
        return zip_path + '.zip', "Model ready for download."
    except Exception as e:
        return None, f"Error saving model: {str(e)}"


def load_example_active_learning():
    """Load the active learning example dataset."""
    return (
        os.path.join(EXAMPLES_DIR, "active_learning", "seed.tsv"),
        os.path.join(EXAMPLES_DIR, "active_learning", "pool.txt"),
        os.path.join(EXAMPLES_DIR, "binary", "dev.tsv"),
        "Binary",
    )


def run_comparison(
    train_file, dev_file, test_file, task_type, selected_models,
    epochs, batch_size, lr, cmp_use_lora, cmp_lora_rank, cmp_lora_alpha,
    progress=gr.Progress(track_tqdm=True),
):
    """Train multiple models on the same data and compare performance + ROC curves."""
    import plotly.graph_objects as go
    from plotly.subplots import make_subplots

    empty = ("", None, None, None, gr.Column(visible=False))
    try:
        if not selected_models or len(selected_models) < 2:
            return ("Select at least 2 models to compare.",) + empty[1:]
        if train_file is None or dev_file is None or test_file is None:
            return ("Upload all 3 data files first.",) + empty[1:]

        epochs = int(epochs)
        batch_size = int(batch_size)

        train_texts, train_labels, n_train = parse_data_file(train_file)
        dev_texts, dev_labels, n_dev = parse_data_file(dev_file)
        test_texts, test_labels, n_test = parse_data_file(test_file)
        num_labels = max(n_train, n_dev, n_test)
        if task_type == "Binary":
            num_labels = 2

        # Only keep these metrics for the table and bar chart
        if task_type == "Binary":
            keep_metrics = {'Accuracy', 'Precision', 'Recall', 'F1'}
        else:
            keep_metrics = {
                'Accuracy', 'F1 Macro', 'F1 Micro',
                'Precision Macro', 'Recall Macro',
            }

        results = []
        roc_data = {}  # model_name -> (true_labels, probabilities)
        log_lines = []

        for i, model_display_name in enumerate(selected_models):
            model_id = FINETUNE_MODELS[model_display_name]
            short_name = model_display_name.split(" (")[0]
            log_lines.append(f"[{i + 1}/{len(selected_models)}] Training {short_name}...")

            try:
                tokenizer = AutoTokenizer.from_pretrained(model_id)
                model = AutoModelForSequenceClassification.from_pretrained(
                    model_id, num_labels=num_labels,
                )

                cmp_lora_active = False
                if cmp_use_lora and PEFT_AVAILABLE:
                    lora_cfg = LoraConfig(
                        task_type=TaskType.SEQ_CLS,
                        r=int(cmp_lora_rank), lora_alpha=int(cmp_lora_alpha),
                        lora_dropout=0.1, bias="none",
                    )
                    model.enable_input_require_grads()
                    model = get_peft_model(model, lora_cfg)
                    cmp_lora_active = True

                train_ds = TextClassificationDataset(train_texts, train_labels, tokenizer, 512)
                dev_ds = TextClassificationDataset(dev_texts, dev_labels, tokenizer, 512)
                test_ds = TextClassificationDataset(test_texts, test_labels, tokenizer, 512)

                output_dir = tempfile.mkdtemp(prefix='conflibert_cmp_')
                best_metric = 'f1' if task_type == 'Binary' else 'f1_macro'

                training_args = TrainingArguments(
                    output_dir=output_dir,
                    num_train_epochs=epochs,
                    per_device_train_batch_size=batch_size,
                    per_device_eval_batch_size=batch_size * 2,
                    learning_rate=lr,
                    weight_decay=0.01,
                    warmup_ratio=0.1,
                    eval_strategy='epoch',
                    save_strategy='epoch',
                    load_best_model_at_end=True,
                    metric_for_best_model=best_metric,
                    greater_is_better=True,
                    logging_steps=50,
                    save_total_limit=1,
                    report_to='none',
                    seed=42,
                )

                trainer = Trainer(
                    model=model,
                    args=training_args,
                    train_dataset=train_ds,
                    eval_dataset=dev_ds,
                    compute_metrics=make_compute_metrics(task_type),
                )

                train_result = trainer.train()

                # Merge LoRA weights before prediction
                if cmp_lora_active and hasattr(trainer.model, 'merge_and_unload'):
                    trainer.model = trainer.model.merge_and_unload()

                # Get predictions for ROC curves
                pred_output = trainer.predict(test_ds)
                logits = pred_output.predictions
                true_labels = pred_output.label_ids
                probs = torch.softmax(torch.tensor(logits), dim=1).numpy()
                roc_data[short_name] = (true_labels, probs)

                # Collect classification metrics only
                test_results = trainer.evaluate(test_ds, metric_key_prefix='test')
                row = {'Model': short_name}
                for k, v in sorted(test_results.items()):
                    if not isinstance(v, (int, float, np.floating, np.integer)):
                        continue
                    name = k.replace('test_', '').replace('_', ' ').title()
                    if name in keep_metrics:
                        row[name] = round(float(v), 4)
                results.append(row)

                runtime = train_result.metrics.get('train_runtime', 0)
                log_lines.append(f"    Done in {runtime:.1f}s")

                del model, trainer, tokenizer, train_ds, dev_ds, test_ds
                gc.collect()
                if torch.cuda.is_available():
                    torch.cuda.empty_cache()

            except Exception as e:
                log_lines.append(f"    Failed: {str(e)}")

        log_lines.append(f"\nComparison complete. {len(results)} models evaluated.")
        log_text = "\n".join(log_lines)

        if not results:
            return log_text, None, None, None, gr.Column(visible=False)

        comparison_df = pd.DataFrame(results)

        # --- Bar chart: classification metrics only ---
        metric_cols = [c for c in comparison_df.columns if c in keep_metrics]
        colors = ['#ff6b35', '#3b82f6', '#10b981', '#8b5cf6', '#f59e0b']
        fig_bar = go.Figure()
        for j, metric in enumerate(metric_cols):
            fig_bar.add_trace(go.Bar(
                name=metric,
                x=comparison_df['Model'],
                y=comparison_df[metric],
                text=comparison_df[metric].apply(
                    lambda x: f'{x:.3f}' if isinstance(x, float) else ''
                ),
                textposition='auto',
                marker_color=colors[j % len(colors)],
            ))
        fig_bar.update_layout(
            barmode='group',
            yaxis_title='Score', yaxis_range=[0, 1.05],
            template='plotly_white',
            legend=dict(orientation='h', yanchor='bottom', y=1.02, xanchor='right', x=1),
            height=400, margin=dict(t=40, b=40),
        )

        # --- ROC curves ---
        model_colors = ['#ff6b35', '#3b82f6', '#10b981', '#8b5cf6',
                        '#f59e0b', '#ec4899', '#06b6d4']
        fig_roc = go.Figure()
        for j, (model_name, (labels, probs)) in enumerate(roc_data.items()):
            color = model_colors[j % len(model_colors)]
            if num_labels == 2:
                fpr, tpr, _ = roc_curve(labels, probs[:, 1])
                roc_auc_val = sk_auc(fpr, tpr)
                fig_roc.add_trace(go.Scatter(
                    x=fpr, y=tpr, mode='lines',
                    name=f'{model_name} (AUC = {roc_auc_val:.3f})',
                    line=dict(color=color, width=2),
                ))
            else:
                # Macro-average ROC for multiclass
                labels_bin = label_binarize(labels, classes=list(range(num_labels)))
                all_fpr = np.linspace(0, 1, 200)
                mean_tpr = np.zeros_like(all_fpr)
                for c in range(num_labels):
                    fpr_c, tpr_c, _ = roc_curve(labels_bin[:, c], probs[:, c])
                    mean_tpr += np.interp(all_fpr, fpr_c, tpr_c)
                mean_tpr /= num_labels
                roc_auc_val = sk_auc(all_fpr, mean_tpr)
                fig_roc.add_trace(go.Scatter(
                    x=all_fpr, y=mean_tpr, mode='lines',
                    name=f'{model_name} (macro AUC = {roc_auc_val:.3f})',
                    line=dict(color=color, width=2),
                ))

        # Diagonal reference line
        fig_roc.add_trace(go.Scatter(
            x=[0, 1], y=[0, 1], mode='lines',
            line=dict(dash='dash', color='#ccc', width=1),
            showlegend=False,
        ))
        fig_roc.update_layout(
            xaxis_title='False Positive Rate',
            yaxis_title='True Positive Rate',
            template='plotly_white',
            legend=dict(orientation='h', yanchor='bottom', y=1.02, xanchor='right', x=1),
            height=400, margin=dict(t=40, b=40),
        )

        return log_text, comparison_df, fig_bar, fig_roc, gr.Column(visible=True)

    except Exception as e:
        return f"Comparison failed: {str(e)}", None, None, None, gr.Column(visible=False)


# ============================================================================
# THEME & CSS
# ============================================================================

utd_orange = gr.themes.Color(
    c50="#fff7f3", c100="#ffead9", c200="#ffd4b3", c300="#ffb380",
    c400="#ff8c52", c500="#ff6b35", c600="#e8551f", c700="#c2410c",
    c800="#9a3412", c900="#7c2d12", c950="#431407",
)

theme = gr.themes.Soft(
    primary_hue=utd_orange,
    secondary_hue="neutral",
    font=[gr.themes.GoogleFont("Inter"), "ui-sans-serif", "system-ui", "sans-serif"],
)

custom_css = """
/* Top accent bar */
.gradio-container::before {
    content: '';
    display: block;
    height: 4px;
    background: linear-gradient(90deg, #ff6b35, #ff9f40, #ff6b35);
    position: fixed;
    top: 0;
    left: 0;
    right: 0;
    z-index: 1000;
}

/* Active tab styling */
.tab-nav button.selected {
    border-bottom-color: #ff6b35 !important;
    color: #ff6b35 !important;
    font-weight: 600 !important;
}

/* Log output - monospace */
.log-output textarea {
    font-family: 'JetBrains Mono', 'Fira Code', 'Consolas', monospace !important;
    font-size: 0.8rem !important;
    line-height: 1.5 !important;
}

/* Dark mode: info callout adjustment */
.dark .info-callout-inner {
    background: rgba(255, 107, 53, 0.1) !important;
    color: #ffead9 !important;
}

/* Clean container width */
.gradio-container {
    max-width: 1200px !important;
}

/* Smooth transitions */
.gradio-container * {
    transition: background-color 0.2s ease, border-color 0.2s ease !important;
}
"""


# ============================================================================
# GRADIO UI
# ============================================================================

with gr.Blocks(theme=theme, css=custom_css, title="ConfliBERT") as demo:

    # ---- HEADER ----
    gr.Markdown(
        "<div style='text-align: center; padding: 1.5rem 0 0.5rem;'>"
        "<h1 style='font-size: 2.5rem; font-weight: 800; margin: 0;'>"
        "<a href='https://eventdata.utdallas.edu/conflibert/' target='_blank' "
        "style='color: #ff6b35; text-decoration: none;'>ConfliBERT</a></h1>"
        "<p style='color: #888; font-size: 0.95rem; margin: 0.25rem 0 0;'>"
        "A Pretrained Language Model for Conflict and Political Violence</p></div>"
    )

    with gr.Tabs():

        # ================================================================
        # HOME TAB
        # ================================================================
        with gr.Tab("Home"):
            gr.Markdown(
                "## Welcome to ConfliBERT\n\n"
                "ConfliBERT is a pretrained language model built specifically for "
                "conflict and political violence text. This application lets you "
                "run inference with ConfliBERT's pretrained models and fine-tune "
                "your own classifiers on custom data. Use the tabs above to get started."
            )

            with gr.Row(equal_height=True):
                with gr.Column():
                    gr.Markdown(
                        "### Inference\n\n"
                        "Run pretrained ConfliBERT models on your text. "
                        "Each task has its own tab with single-text analysis "
                        "and CSV batch processing.\n\n"
                        "**Named Entity Recognition**\n"
                        "Identify persons, organizations, locations, weapons, "
                        "and other entities in text. Results are color-coded "
                        "by entity type.\n\n"
                        "**Binary Classification**\n"
                        "Determine whether text is related to conflict, violence, "
                        "or politics (positive) or not (negative). You can also "
                        "load a custom fine-tuned model here.\n\n"
                        "**Multilabel Classification**\n"
                        "Score text against four event categories: Armed Assault, "
                        "Bombing/Explosion, Kidnapping, and Other. Each category "
                        "is scored independently.\n\n"
                        "**Question Answering**\n"
                        "Provide a context passage and ask a question. The model "
                        "extracts the most relevant answer span from the text."
                    )
                with gr.Column():
                    gr.Markdown(
                        "### Fine-tuning\n\n"
                        "Train your own binary or multiclass text classifier "
                        "on custom labeled data, all within the browser.\n\n"
                        "**Workflow:**\n"
                        "1. Upload your training, validation, and test data as "
                        "TSV files (or load a built-in example dataset)\n"
                        "2. Pick a base model: ConfliBERT, BERT, RoBERTa, "
                        "ModernBERT, DeBERTa, or DistilBERT\n"
                        "3. Configure training parameters (sensible defaults "
                        "are provided)\n"
                        "4. Train and watch progress in real time\n"
                        "5. Review test-set metrics (accuracy, precision, "
                        "recall, F1)\n"
                        "6. Try your model on new text immediately\n"
                        "7. Run batch predictions on a CSV\n"
                        "8. Save the model and load it later in the "
                        "Classification tab\n\n"
                        "**Advanced features:**\n"
                        "- **LoRA / QLoRA** for parameter-efficient training "
                        "(lower VRAM, faster)\n"
                        "- **Active Learning** tab for iterative labeling "
                        "with uncertainty sampling\n"
                        "- Early stopping with configurable patience\n"
                        "- Learning rate schedulers (linear, cosine, constant)\n"
                        "- Mixed precision training (FP16 on CUDA GPUs)\n"
                        "- Gradient accumulation for larger effective batch sizes\n"
                        "- Weight decay regularization"
                    )

            gr.Markdown(
                f"---\n\n"
                f"**Your system:** {get_system_info()}"
            )

            gr.Markdown(
                "**Citation:** Brandt, P.T., Alsarra, S., D'Orazio, V., "
                "Heintze, D., Khan, L., Meher, S., Osorio, J. and Sianan, M., "
                "2025. Extractive versus Generative Language Models for Political "
                "Conflict Text Classification. *Political Analysis*, pp.1-29."
            )

        # ================================================================
        # NER TAB
        # ================================================================
        with gr.Tab("Named Entity Recognition"):
            gr.Markdown(info_callout(
                "Identify entities in text such as **persons**, **organizations**, "
                "**locations**, **weapons**, and more. Results are color-coded by type."
            ))
            with gr.Row(equal_height=True):
                with gr.Column():
                    ner_input = gr.Textbox(
                        lines=6,
                        placeholder="Paste or type text to analyze for entities...",
                        label="Input Text",
                    )
                    ner_btn = gr.Button("Analyze Entities", variant="primary")
                with gr.Column():
                    ner_output = gr.HTML(label="Results")

            with gr.Accordion("Batch Processing (CSV)", open=False):
                gr.Markdown(
                    "Upload a CSV file with a `text` column to process "
                    "multiple texts at once."
                )
                with gr.Row():
                    ner_csv_in = gr.File(
                        label="Upload CSV", file_types=[".csv"],
                    )
                    ner_csv_out = gr.File(label="Download Results")
                ner_csv_btn = gr.Button("Process CSV", variant="secondary")

        # ================================================================
        # BINARY CLASSIFICATION TAB
        # ================================================================
        with gr.Tab("Binary Classification"):
            gr.Markdown(info_callout(
                "Classify text as **conflict-related** (positive) or "
                "**not conflict-related** (negative). Uses the pretrained ConfliBERT "
                "binary classifier by default, or load your own finetuned model below."
            ))

            custom_clf_model = gr.State(None)
            custom_clf_tokenizer = gr.State(None)

            with gr.Row(equal_height=True):
                with gr.Column():
                    clf_input = gr.Textbox(
                        lines=6,
                        placeholder="Paste or type text to classify...",
                        label="Input Text",
                    )
                    clf_btn = gr.Button("Classify", variant="primary")
                with gr.Column():
                    clf_output = gr.HTML(label="Results")

            with gr.Accordion("Batch Processing (CSV)", open=False):
                gr.Markdown("Upload a CSV file with a `text` column.")
                with gr.Row():
                    clf_csv_in = gr.File(label="Upload CSV", file_types=[".csv"])
                    clf_csv_out = gr.File(label="Download Results")
                clf_csv_btn = gr.Button("Process CSV", variant="secondary")

            with gr.Accordion("Load Custom Model", open=False):
                gr.Markdown(
                    "Load a finetuned classification model from a local directory "
                    "to use instead of the default pretrained classifier."
                )
                clf_model_path = gr.Textbox(
                    label="Model directory path",
                    placeholder="e.g., ./finetuned_model",
                )
                with gr.Row():
                    clf_load_btn = gr.Button("Load Model", variant="secondary")
                    clf_reset_btn = gr.Button(
                        "Reset to Pretrained", variant="secondary",
                    )
                clf_status = gr.Markdown("")

        # ================================================================
        # MULTILABEL CLASSIFICATION TAB
        # ================================================================
        with gr.Tab("Multilabel Classification"):
            gr.Markdown(info_callout(
                "Identify multiple event types in text. Each category is scored "
                "independently: **Armed Assault**, **Bombing/Explosion**, "
                "**Kidnapping**, **Other**. Categories above 50% confidence "
                "are highlighted. Load a custom finetuned model below."
            ))

            custom_multi_model = gr.State(None)
            custom_multi_tokenizer = gr.State(None)

            with gr.Row(equal_height=True):
                with gr.Column():
                    multi_input = gr.Textbox(
                        lines=6,
                        placeholder="Paste or type text to classify...",
                        label="Input Text",
                    )
                    multi_btn = gr.Button("Classify", variant="primary")
                with gr.Column():
                    multi_output = gr.HTML(label="Results")

            with gr.Accordion("Batch Processing (CSV)", open=False):
                gr.Markdown("Upload a CSV file with a `text` column.")
                with gr.Row():
                    multi_csv_in = gr.File(label="Upload CSV", file_types=[".csv"])
                    multi_csv_out = gr.File(label="Download Results")
                multi_csv_btn = gr.Button("Process CSV", variant="secondary")

            with gr.Accordion("Load Custom Model", open=False):
                gr.Markdown(
                    "Load a finetuned multiclass model from a local directory "
                    "to use instead of the default pretrained classifier."
                )
                multi_model_path = gr.Textbox(
                    label="Model directory path",
                    placeholder="e.g., ./finetuned_model",
                )
                with gr.Row():
                    multi_load_btn = gr.Button("Load Model", variant="secondary")
                    multi_reset_btn = gr.Button(
                        "Reset to Pretrained", variant="secondary",
                    )
                multi_status = gr.Markdown("")

        # ================================================================
        # QUESTION ANSWERING TAB
        # ================================================================
        with gr.Tab("Question Answering"):
            gr.Markdown(info_callout(
                "Extract answers from a context passage. Provide a paragraph of "
                "text and ask a question about it. The model will highlight the "
                "most relevant span."
            ))
            with gr.Row(equal_height=True):
                with gr.Column():
                    qa_context = gr.Textbox(
                        lines=6,
                        placeholder="Paste the context passage here...",
                        label="Context",
                    )
                    qa_question = gr.Textbox(
                        lines=2,
                        placeholder="What would you like to know?",
                        label="Question",
                    )
                    qa_btn = gr.Button("Get Answer", variant="primary")
                with gr.Column():
                    qa_output = gr.HTML(label="Answer")

            with gr.Accordion("Batch Processing (CSV)", open=False):
                gr.Markdown(
                    "Upload a CSV file with `context` and `question` columns "
                    "to process multiple questions at once."
                )
                with gr.Row():
                    qa_csv_in = gr.File(
                        label="Upload CSV", file_types=[".csv"],
                    )
                    qa_csv_out = gr.File(label="Download Results")
                qa_csv_btn = gr.Button("Process CSV", variant="secondary")

        # ================================================================
        # FINE-TUNE TAB
        # ================================================================
        with gr.Tab("Fine-tune"):
            gr.Markdown(info_callout(
                "Fine-tune a binary or multiclass classifier on your own data. "
                "Upload labeled TSV files, pick a base model, and train. "
                "Or compare multiple models head-to-head on the same dataset."
            ))

            # -- Data --
            gr.Markdown("### Data")
            gr.Markdown(
                "TSV files, no header, format: `text[TAB]label` "
                "(binary: 0/1, multiclass: 0, 1, 2, ...)"
            )
            with gr.Row():
                ft_ex_binary_btn = gr.Button(
                    "Load Example: Binary", variant="secondary", size="sm",
                )
                ft_ex_multi_btn = gr.Button(
                    "Load Example: Multiclass (4 classes)", variant="secondary", size="sm",
                )
            with gr.Row():
                ft_train_file = gr.File(
                    label="Train", file_types=[".tsv", ".csv", ".txt"],
                )
                ft_dev_file = gr.File(
                    label="Validation", file_types=[".tsv", ".csv", ".txt"],
                )
                ft_test_file = gr.File(
                    label="Test", file_types=[".tsv", ".csv", ".txt"],
                )

            # -- Configuration --
            gr.Markdown("### Configuration")
            with gr.Row():
                ft_task = gr.Radio(
                    ["Binary", "Multiclass"],
                    label="Task Type", value="Binary",
                )
                ft_model = gr.Dropdown(
                    choices=list(FINETUNE_MODELS.keys()),
                    label="Base Model",
                    value=list(FINETUNE_MODELS.keys())[0],
                )
            with gr.Row():
                ft_epochs = gr.Number(
                    label="Epochs", value=3, minimum=1, maximum=100, precision=0,
                )
                ft_batch = gr.Number(
                    label="Batch Size", value=8, minimum=1, maximum=128, precision=0,
                )
                ft_lr = gr.Number(
                    label="Learning Rate", value=2e-5, minimum=1e-7, maximum=1e-2,
                )

            with gr.Accordion("Advanced Settings", open=False):
                with gr.Row():
                    ft_weight_decay = gr.Number(
                        label="Weight Decay", value=0.01, minimum=0, maximum=1,
                    )
                    ft_warmup = gr.Number(
                        label="Warmup Ratio", value=0.1, minimum=0, maximum=0.5,
                    )
                    ft_max_len = gr.Number(
                        label="Max Sequence Length", value=512,
                        minimum=32, maximum=8192, precision=0,
                    )
                with gr.Row():
                    ft_grad_accum = gr.Number(
                        label="Gradient Accumulation", value=1,
                        minimum=1, maximum=64, precision=0,
                    )
                    ft_fp16 = gr.Checkbox(
                        label="Mixed Precision (FP16)", value=False,
                    )
                    ft_patience = gr.Number(
                        label="Early Stopping Patience", value=3,
                        minimum=0, maximum=20, precision=0,
                    )
                ft_scheduler = gr.Dropdown(
                    ["linear", "cosine", "constant", "constant_with_warmup"],
                    label="LR Scheduler", value="linear",
                )
                gr.Markdown("**Parameter-Efficient Fine-Tuning (PEFT)**")
                with gr.Row():
                    ft_use_lora = gr.Checkbox(
                        label="Use LoRA", value=False,
                    )
                    ft_lora_rank = gr.Number(
                        label="LoRA Rank (r)", value=8,
                        minimum=1, maximum=256, precision=0,
                    )
                    ft_lora_alpha = gr.Number(
                        label="LoRA Alpha", value=16,
                        minimum=1, maximum=512, precision=0,
                    )
                    ft_use_qlora = gr.Checkbox(
                        label="QLoRA (4-bit, CUDA only)", value=False,
                    )

            # -- Train --
            ft_train_btn = gr.Button(
                "Start Training", variant="primary", size="lg",
            )

            # State for the trained model
            ft_model_state = gr.State(None)
            ft_tokenizer_state = gr.State(None)
            ft_num_labels_state = gr.State(None)

            with gr.Accordion("Training Log", open=False) as ft_log_accordion:
                ft_log = gr.Textbox(
                    lines=12, interactive=False, elem_classes="log-output",
                    show_label=False,
                )

            # -- Results + Try Model (hidden until training completes) --
            with gr.Column(visible=False) as ft_results_col:
                gr.Markdown("### Results")
                with gr.Row(equal_height=True):
                    with gr.Column(scale=2):
                        ft_metrics = gr.Dataframe(
                            label="Test Set Metrics",
                            headers=["Metric", "Score"],
                            interactive=False,
                        )
                    with gr.Column(scale=3):
                        gr.Markdown("**Try your model**")
                        ft_try_input = gr.Textbox(
                            lines=2, label="Input Text",
                            placeholder="Type text to classify...",
                        )
                        with gr.Row():
                            ft_try_btn = gr.Button("Predict", variant="primary")
                        ft_try_output = gr.HTML(label="Prediction")

            # -- Save + Batch (hidden until training completes) --
            with gr.Column(visible=False) as ft_actions_col:
                with gr.Row(equal_height=True):
                    with gr.Column():
                        gr.Markdown("**Download model**")
                        ft_save_btn = gr.Button("Prepare Download", variant="secondary")
                        ft_save_file = gr.File(label="Download Model (.zip)")
                        ft_save_status = gr.Markdown("")
                    with gr.Column():
                        gr.Markdown("**Batch predictions**")
                        ft_batch_in = gr.File(
                            label="Upload CSV (needs 'text' column)",
                            file_types=[".csv"],
                        )
                        ft_batch_btn = gr.Button(
                            "Run Predictions", variant="secondary",
                        )
                        ft_batch_out = gr.File(label="Download Results")

            # -- Compare Models --
            gr.Markdown("---")
            with gr.Accordion("Compare Multiple Models", open=False):
                gr.Markdown(
                    "Train the same dataset on different base models and compare "
                    "performance side by side. Uses the data and task type above."
                )
                cmp_models = gr.CheckboxGroup(
                    choices=list(FINETUNE_MODELS.keys()),
                    label="Select models to compare (pick 2 or more)",
                )
                with gr.Row():
                    cmp_epochs = gr.Number(label="Epochs", value=3, minimum=1, precision=0)
                    cmp_batch = gr.Number(label="Batch Size", value=8, minimum=1, precision=0)
                    cmp_lr = gr.Number(label="Learning Rate", value=2e-5, minimum=1e-7)
                with gr.Row():
                    cmp_use_lora = gr.Checkbox(label="Use LoRA", value=False)
                    cmp_lora_rank = gr.Number(label="LoRA Rank", value=8, minimum=1, maximum=256, precision=0)
                    cmp_lora_alpha = gr.Number(label="LoRA Alpha", value=16, minimum=1, maximum=512, precision=0)
                cmp_btn = gr.Button("Compare Models", variant="primary")
                cmp_log = gr.Textbox(
                    label="Comparison Log", lines=8,
                    interactive=False, elem_classes="log-output",
                )
                with gr.Column(visible=False) as cmp_results_col:
                    cmp_table = gr.Dataframe(
                        label="Comparison Results", interactive=False,
                    )
                    cmp_plot = gr.Plot(label="Metrics Comparison")
                    cmp_roc = gr.Plot(label="ROC Curves")

        # ================================================================
        # ACTIVE LEARNING TAB
        # ================================================================
        with gr.Tab("Active Learning"):
            gr.Markdown(info_callout(
                "**Active learning** iteratively selects the most uncertain "
                "samples from an unlabeled pool for you to label, then retrains. "
                "This lets you build a strong classifier with far fewer labels."
            ))

            # -- Data --
            gr.Markdown("### Data")
            gr.Markdown(
                "**Seed file** — small labeled set (TSV, `text[TAB]label`).  \n"
                "**Pool file** — unlabeled texts (one per line, or CSV with `text` column).  \n"
                "**Dev file** *(optional)* — held-out labeled set to track metrics."
            )
            al_ex_btn = gr.Button(
                "Load Example: Binary Active Learning",
                variant="secondary", size="sm",
            )
            with gr.Row():
                al_seed_file = gr.File(
                    label="Labeled Seed (TSV)",
                    file_types=[".tsv", ".csv", ".txt"],
                )
                al_pool_file = gr.File(
                    label="Unlabeled Pool",
                    file_types=[".tsv", ".csv", ".txt"],
                )
                al_dev_file = gr.File(
                    label="Dev / Validation (optional)",
                    file_types=[".tsv", ".csv", ".txt"],
                )

            # -- Configuration --
            gr.Markdown("### Configuration")
            with gr.Row():
                al_task = gr.Radio(
                    ["Binary", "Multiclass"],
                    label="Task Type", value="Binary",
                )
                al_model_dd = gr.Dropdown(
                    choices=list(FINETUNE_MODELS.keys()),
                    label="Base Model",
                    value=list(FINETUNE_MODELS.keys())[0],
                )
            with gr.Row():
                al_strategy = gr.Dropdown(
                    ["entropy", "margin", "least_confidence"],
                    label="Query Strategy", value="entropy",
                )
                al_query_size = gr.Number(
                    label="Samples per Round", value=20,
                    minimum=1, maximum=500, precision=0,
                )
            with gr.Row():
                al_epochs = gr.Number(
                    label="Epochs per Round", value=3,
                    minimum=1, maximum=50, precision=0,
                )
                al_batch_size = gr.Number(
                    label="Batch Size", value=8,
                    minimum=1, maximum=128, precision=0,
                )
                al_lr = gr.Number(
                    label="Learning Rate", value=2e-5,
                    minimum=1e-7, maximum=1e-2,
                )
            with gr.Accordion("Advanced", open=False):
                with gr.Row():
                    al_max_len = gr.Number(
                        label="Max Sequence Length", value=512,
                        minimum=32, maximum=8192, precision=0,
                    )
                    al_use_lora = gr.Checkbox(label="Use LoRA", value=False)
                    al_lora_rank = gr.Number(
                        label="LoRA Rank", value=8,
                        minimum=1, maximum=256, precision=0,
                    )
                    al_lora_alpha = gr.Number(
                        label="LoRA Alpha", value=16,
                        minimum=1, maximum=512, precision=0,
                    )

            al_init_btn = gr.Button(
                "Initialize Active Learning", variant="primary", size="lg",
            )

            # -- State --
            al_state = gr.State({})
            al_model_state = gr.State(None)
            al_tokenizer_state = gr.State(None)

            with gr.Accordion("Log", open=False):
                al_log = gr.Textbox(
                    lines=12, interactive=False, elem_classes="log-output",
                    show_label=False,
                )

            # -- Annotation panel (hidden until init) --
            with gr.Column(visible=False) as al_annotation_col:
                gr.Markdown("### Label These Samples")
                gr.Markdown(
                    "Fill in the **Label** column with integer class labels "
                    "(e.g. 0 or 1 for binary). Then click **Submit**."
                )
                al_annotation_df = gr.Dataframe(
                    headers=["Text", "Label"],
                    interactive=True,
                )
                with gr.Row():
                    al_submit_btn = gr.Button(
                        "Submit Labels & Next Round",
                        variant="primary",
                    )

                al_chart = gr.Plot(label="Metrics Across Rounds")

                gr.Markdown("### Download Model")
                with gr.Row():
                    al_save_btn = gr.Button("Prepare Download", variant="secondary")
                    al_save_file = gr.File(label="Download Model (.zip)")
                    al_save_status = gr.Markdown("")

    # ---- FOOTER ----
    gr.Markdown(
        "<div style='text-align: center; padding: 1rem 0; margin-top: 0.5rem; "
        "border-top: 1px solid #e5e7eb;'>"
        "<p style='color: #888; font-size: 0.85rem; margin: 0;'>"
        "Developed by "
        "<a href='http://shreyasmeher.com' target='_blank' "
        "style='color: #ff6b35; text-decoration: none;'>Shreyas Meher</a>"
        "</p>"
        "<p style='color: #999; font-size: 0.75rem; margin: 0.5rem 0 0; "
        "max-width: 700px; margin-left: auto; margin-right: auto; line-height: 1.4;'>"
        "If you use ConfliBERT in your research, please cite:<br>"
        "<em>Brandt, P.T., Alsarra, S., D'Orazio, V., Heintze, D., Khan, L., "
        "Meher, S., Osorio, J. and Sianan, M., 2025. Extractive versus Generative "
        "Language Models for Political Conflict Text Classification. "
        "Political Analysis, pp.1&ndash;29.</em>"
        "</p></div>"
    )

    # ====================================================================
    # EVENT HANDLERS
    # ====================================================================

    # NER
    ner_btn.click(
        fn=named_entity_recognition, inputs=[ner_input], outputs=[ner_output],
    )
    ner_csv_btn.click(
        fn=process_csv_ner, inputs=[ner_csv_in], outputs=[ner_csv_out],
    )

    # Binary Classification
    clf_btn.click(
        fn=text_classification,
        inputs=[clf_input, custom_clf_model, custom_clf_tokenizer],
        outputs=[clf_output],
    )
    clf_csv_btn.click(
        fn=process_csv_binary,
        inputs=[clf_csv_in, custom_clf_model, custom_clf_tokenizer],
        outputs=[clf_csv_out],
    )
    clf_load_btn.click(
        fn=load_custom_model,
        inputs=[clf_model_path],
        outputs=[custom_clf_model, custom_clf_tokenizer, clf_status],
    )
    clf_reset_btn.click(
        fn=reset_custom_model,
        outputs=[custom_clf_model, custom_clf_tokenizer, clf_status],
    )

    # Multilabel Classification
    multi_btn.click(
        fn=multilabel_classification,
        inputs=[multi_input, custom_multi_model, custom_multi_tokenizer],
        outputs=[multi_output],
    )
    multi_csv_btn.click(
        fn=process_csv_multilabel, inputs=[multi_csv_in], outputs=[multi_csv_out],
    )
    multi_load_btn.click(
        fn=load_custom_model,
        inputs=[multi_model_path],
        outputs=[custom_multi_model, custom_multi_tokenizer, multi_status],
    )
    multi_reset_btn.click(
        fn=reset_custom_model,
        outputs=[custom_multi_model, custom_multi_tokenizer, multi_status],
    )

    # Question Answering
    qa_btn.click(
        fn=question_answering,
        inputs=[qa_context, qa_question],
        outputs=[qa_output],
    )
    qa_csv_btn.click(
        fn=process_csv_qa, inputs=[qa_csv_in], outputs=[qa_csv_out],
    )

    # Fine-tuning: example dataset loaders
    ft_ex_binary_btn.click(
        fn=load_example_binary,
        outputs=[ft_train_file, ft_dev_file, ft_test_file, ft_task],
    )
    ft_ex_multi_btn.click(
        fn=load_example_multiclass,
        outputs=[ft_train_file, ft_dev_file, ft_test_file, ft_task],
    )

    # Fine-tuning: training
    ft_train_btn.click(
        fn=run_finetuning,
        inputs=[
            ft_train_file, ft_dev_file, ft_test_file,
            ft_task, ft_model,
            ft_epochs, ft_batch, ft_lr,
            ft_weight_decay, ft_warmup, ft_max_len,
            ft_grad_accum, ft_fp16, ft_patience, ft_scheduler,
            ft_use_lora, ft_lora_rank, ft_lora_alpha, ft_use_qlora,
        ],
        outputs=[
            ft_log, ft_metrics,
            ft_model_state, ft_tokenizer_state, ft_num_labels_state,
            ft_results_col, ft_actions_col,
        ],
        concurrency_limit=1,
    )

    # Try finetuned model
    ft_try_btn.click(
        fn=predict_finetuned,
        inputs=[ft_try_input, ft_model_state, ft_tokenizer_state, ft_num_labels_state],
        outputs=[ft_try_output],
    )

    # Save finetuned model
    ft_save_btn.click(
        fn=save_finetuned_model,
        inputs=[ft_model_state, ft_tokenizer_state],
        outputs=[ft_save_file, ft_save_status],
    )

    # Batch predictions with finetuned model
    ft_batch_btn.click(
        fn=batch_predict_finetuned,
        inputs=[ft_batch_in, ft_model_state, ft_tokenizer_state, ft_num_labels_state],
        outputs=[ft_batch_out],
    )

    # Active Learning: example loader
    al_ex_btn.click(
        fn=load_example_active_learning,
        outputs=[al_seed_file, al_pool_file, al_dev_file, al_task],
    )

    # Active Learning
    al_init_btn.click(
        fn=al_initialize,
        inputs=[
            al_seed_file, al_pool_file, al_dev_file,
            al_task, al_model_dd, al_strategy, al_query_size,
            al_epochs, al_batch_size, al_lr, al_max_len,
            al_use_lora, al_lora_rank, al_lora_alpha,
        ],
        outputs=[
            al_state, al_model_state, al_tokenizer_state,
            al_annotation_df, al_log, al_chart,
            al_annotation_col,
        ],
        concurrency_limit=1,
    )

    al_submit_btn.click(
        fn=al_submit_and_continue,
        inputs=[
            al_annotation_df, al_state, al_model_state, al_tokenizer_state,
            al_log,
        ],
        outputs=[
            al_state, al_model_state, al_tokenizer_state,
            al_annotation_df, al_log, al_chart,
        ],
        concurrency_limit=1,
    )

    al_save_btn.click(
        fn=al_save_model,
        inputs=[al_model_state, al_tokenizer_state],
        outputs=[al_save_file, al_save_status],
    )

    # Model comparison
    cmp_btn.click(
        fn=run_comparison,
        inputs=[
            ft_train_file, ft_dev_file, ft_test_file,
            ft_task, cmp_models, cmp_epochs, cmp_batch, cmp_lr,
            cmp_use_lora, cmp_lora_rank, cmp_lora_alpha,
        ],
        outputs=[cmp_log, cmp_table, cmp_plot, cmp_roc, cmp_results_col],
        concurrency_limit=1,
    )


# ============================================================================
# LAUNCH
# ============================================================================

demo.launch(share=True, ssr_mode=False)