app.py

# -*- coding: utf-8 -*-
"""
Mahoun — Legal AI (RAG + Training + Metrics) for HF Spaces / Gradio 5
- سازگار با Gradio 5.x و Transformers >= 4.44
- TrainingArguments ایمن با عقب‌سازگاری (safe_training_args)
- RAG با ChromaDB + ایندکس‌سازی JSONL قوانین
- متریک‌ها: ROUGE-L (seq2seq) و F1 ساده (causal)
- ماسک پدینگ روی labels در معماری علّی
- Progress به‌صورت DI: progress=gr.Progress(track_tqdm=True)

ساختار ورودی دیتاست آموزش:
  JSONL با کلیدهای "input" و "output"

ساختار ورودی قوانین برای ایندکس:
  JSONL با کلیدهای (پیش‌فرض) "article_id" و "text"
"""

from __future__ import annotations
import os, sys, json, warnings
from dataclasses import dataclass, field
from pathlib import Path
from typing import List, Dict, Optional, Tuple

import numpy as np
import torch
from torch.utils.data import Dataset
from sklearn.model_selection import train_test_split

import gradio as gr
from packaging import version

import transformers as tf
from transformers import (
    AutoTokenizer,
    AutoModelForSeq2SeqLM,
    AutoModelForCausalLM,
    Trainer,
    TrainingArguments,
    EarlyStoppingCallback,
    DataCollatorForSeq2Seq,
)

# RAG stack
import chromadb
from sentence_transformers import SentenceTransformer

# Optional metrics
try:
    from evaluate import load as eval_load
except Exception:
    eval_load = None

warnings.filterwarnings("ignore")

# ==========================
# Config
# ==========================
@dataclass
class ModelConfig:
    model_name: str = "google/mt5-base"
    architecture: str = "seq2seq"  # "seq2seq" | "causal"
    max_input_length: int = 1024
    max_target_length: int = 512
    max_new_tokens: int = 384
    temperature: float = 0.7
    top_p: float = 0.9
    num_beams: int = 4
    gradient_checkpointing: bool = True

@dataclass
class RAGConfig:
    embedding_model: str = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
    persist_dir: str = "./chroma_db"
    collection: str = "legal_articles"
    top_k: int = 5
    similarity_threshold: float = 0.66   # 0..1
    context_char_limit: int = 300
    enable: bool = True

@dataclass
class TrainConfig:
    output_dir: str = "./mahoon_model"
    seed: int = 42
    test_size: float = 0.1
    epochs: int = 3
    batch_size: int = 2
    grad_accum: int = 2
    lr: float = 3e-5
    use_bf16: bool = True
    weight_decay: float = 0.01
    warmup_ratio: float = 0.05
    logging_steps: int = 50
    eval_strategy: str = "epoch"   # "steps" | "epoch"
    save_strategy: str = "epoch"
    save_total_limit: int = 2
    report_to: str = "none"        # "none" | "wandb"
    max_grad_norm: float = 1.0

@dataclass
class SystemConfig:
    model: ModelConfig = field(default_factory=ModelConfig)
    rag: RAGConfig = field(default_factory=RAGConfig)
    train: TrainConfig = field(default_factory=TrainConfig)

# ==========================
# Utils
# ==========================
def set_seed_all(seed: int = 42):
    import random
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(seed)

def log_deps():
    try:
        import accelerate, datasets
        print("[deps]",
              f"python={sys.version.split()[0]}",
              f"transformers={tf.__version__}",
              f"accelerate={accelerate.__version__}",
              f"datasets={datasets.__version__}",
              f"gradio={gr.__version__}",
              flush=True)
    except Exception as e:
        print("[deps] warn:", e, flush=True)

def bf16_supported():
    return torch.cuda.is_available() and hasattr(torch.cuda, "is_bf16_supported") and torch.cuda.is_bf16_supported()

def safe_training_args(**kwargs):
    """
    Wrapper برای سازگاری با نسخه‌های قدیمی‌تر Transformers (قبل از 4.4):
    - evaluation_strategy -> evaluate_during_training
    - حذف کلیدهای جدید که ممکن است ناشناخته باشند
    """
    tf_ver = version.parse(tf.__version__)
    k = dict(kwargs)
    if tf_ver < version.parse("4.4.0"):
        eval_strat = k.pop("evaluation_strategy", None)
        k["evaluate_during_training"] = bool(eval_strat and str(eval_strat).lower() != "no")
        for rm in ["save_strategy","load_best_model_at_end","metric_for_best_model",
                   "greater_is_better","predict_with_generate","generation_max_length",
                   "generation_num_beams","report_to","max_grad_norm"]:
            k.pop(rm, None)
    return TrainingArguments(**k)

# ==========================
# RAG
# ==========================
class LegalRAG:
    def __init__(self, cfg: RAGConfig):
        self.cfg = cfg
        self.client = None
        self.collection = None
        self.embedder: Optional[SentenceTransformer] = None

    def init(self):
        Path(self.cfg.persist_dir).mkdir(parents=True, exist_ok=True)
        self.client = chromadb.PersistentClient(path=self.cfg.persist_dir)
        try:
            self.collection = self.client.get_or_create_collection(self.cfg.collection)
        except Exception:
            try:
                self.collection = self.client.get_collection(self.cfg.collection)
            except Exception:
                self.collection = self.client.create_collection(self.cfg.collection)
        self.embedder = SentenceTransformer(self.cfg.embedding_model)

    def index_jsonl(self, jsonl_path: str, id_key="article_id", text_key="text"):
        """ایندکس‌سازی اولیه قوانین از JSONL: هر خط یک شیء {article_id, text, ...}."""
        if not self.collection or not self.embedder:
            self.init()
        ids, docs, metas = [], [], []
        with open(jsonl_path, "r", encoding="utf-8") as f:
            for i, line in enumerate(f):
                s = line.strip()
                if not s:
                    continue
                try:
                    obj = json.loads(s)
                except:
                    continue
                aid = str(obj.get(id_key, f"auto_{i}"))
                txt = str(obj.get(text_key, "")).strip()
                if not txt:
                    continue
                ids.append(aid)
                docs.append(txt)
                metas.append({"article_id": aid})
        if not ids:
            return "هیچ سندی برای ایندکس پیدا نشد."
        self.collection.upsert(ids=ids, documents=docs, metadatas=metas)
        return f"✅ {len(ids)} سند قانونی ایندکس شد."

    def retrieve(self, query: str) -> List[Dict]:
        if not self.collection:
            return []
        try:
            res = self.collection.query(
                query_texts=[query],
                n_results=self.cfg.top_k,
                include=["documents","metadatas","distances"],
            )
            out = []
            docs = res.get("documents", [[]])[0]
            metas = res.get("metadatas", [[]])[0]
            dists = res.get("distances", [[1.0]])[0]
            for i, (doc, meta, dist) in enumerate(zip(docs, metas, dists)):
                sim = 1.0 - float(dist)
                if sim >= self.cfg.similarity_threshold:
                    out.append({
                        "article_id": (meta or {}).get("article_id", f"unk_{i}"),
                        "text": doc,
                        "similarity": sim,
                    })
            return out
        except Exception:
            return []

    def build_context(self, arts: List[Dict]) -> str:
        if not arts:
            return ""
        bullets = [f"• ماده {a['article_id']}: {a['text'][:self.cfg.context_char_limit]}..." for a in arts]
        return "مواد مرتبط:\n" + "\n".join(bullets)

# ==========================
# Loader + Generator
# ==========================
class ModelLoader:
    def __init__(self, mcfg: ModelConfig):
        self.cfg = mcfg
        self.tokenizer = None
        self.model = None

    def load(self):
        self.tokenizer = AutoTokenizer.from_pretrained(self.cfg.model_name)
        # dtype انتخاب هوشمند
        use_bf16 = bf16_supported() and self.cfg.gradient_checkpointing
        dtype = torch.bfloat16 if use_bf16 else (torch.float16 if torch.cuda.is_available() else None)
        model_kwargs = {"torch_dtype": dtype}
        if torch.cuda.is_available():
            model_kwargs["device_map"] = "auto"

        if self.cfg.architecture == "seq2seq":
            self.model = AutoModelForSeq2SeqLM.from_pretrained(self.cfg.model_name, **model_kwargs)
        elif self.cfg.architecture == "causal":
            self.model = AutoModelForCausalLM.from_pretrained(self.cfg.model_name, **model_kwargs)
            if self.tokenizer.pad_token is None and hasattr(self.tokenizer, "eos_token"):
                self.tokenizer.pad_token = self.tokenizer.eos_token
        else:
            raise ValueError("Unsupported architecture")

        if self.cfg.gradient_checkpointing and hasattr(self.model, "gradient_checkpointing_enable"):
            try:
                self.model.gradient_checkpointing_enable()
            except Exception:
                pass
        return self

class Generator:
    def __init__(self, loader: ModelLoader, mcfg: ModelConfig):
        self.tk = loader.tokenizer
        self.model = loader.model
        self.cfg = mcfg

    def generate(self, question: str, context: str = "") -> str:
        if self.cfg.architecture == "seq2seq":
            inp = f"{context}\nسوال: {question}" if context else f"سوال: {question}"
            enc = self.tk(inp, return_tensors="pt", truncation=True, max_length=self.cfg.max_input_length)
            enc = {k: v.to(self.model.device) for k,v in enc.items()}
            out = self.model.generate(
                **enc,
                max_length=self.cfg.max_target_length,
                num_beams=self.cfg.num_beams,
                early_stopping=True,
            )
        else:
            prompt = f"{context}\nسوال: {question}\nپاسخ:" if context else f"سوال: {question}\nپاسخ:"
            enc = self.tk(prompt, return_tensors="pt", truncation=True, max_length=self.cfg.max_input_length)
            enc = {k: v.to(self.model.device) for k,v in enc.items()}
            out = self.model.generate(
                **enc,
                max_new_tokens=self.cfg.max_new_tokens,
                do_sample=True,
                temperature=self.cfg.temperature,
                top_p=self.cfg.top_p,
                pad_token_id=self.tk.pad_token_id or self.tk.eos_token_id,
            )
        return self.tk.decode(out[0], skip_special_tokens=True)

# ==========================
# Datasets
# ==========================
class Seq2SeqJSONLDataset(Dataset):
    def __init__(self, data: List[Dict], tokenizer, max_inp: int, max_tgt: int, rag: Optional[LegalRAG] = None, enhance_every:int = 10):
        self.tk = tokenizer
        self.max_inp = max_inp
        self.max_tgt = max_tgt
        self.items = []
        for i, ex in enumerate(data):
            src = str(ex.get("input", "")).strip()
            tgt = str(ex.get("output", "")).strip()
            if not src or not tgt:
                continue
            inp = src
            if rag and i % enhance_every == 0:
                arts = rag.retrieve(src)
                ctx = rag.build_context(arts)
                if ctx:
                    inp = f"<CONTEXT>{ctx}</CONTEXT>\n<QUESTION>{src}</QUESTION>"
            self.items.append((inp, tgt))

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

    def __getitem__(self, idx):
        inp, tgt = self.items[idx]
        model_inputs = self.tk(inp, max_length=self.max_inp, padding="max_length", truncation=True)
        labels = self.tk(text_target=tgt, max_length=self.max_tgt, padding="max_length", truncation=True)
        model_inputs["labels"] = labels["input_ids"]
        return model_inputs

class CausalJSONLDataset(Dataset):
    def __init__(self, data: List[Dict], tokenizer, max_inp: int, rag: Optional[LegalRAG] = None, enhance_every:int = 10):
        self.tk = tokenizer
        self.max_inp = max_inp
        self.items = []
        for i, ex in enumerate(data):
            src = str(ex.get("input", "")).strip()
            tgt = str(ex.get("output", "")).strip()
            if not src or not tgt:
                continue
            ctx = ""
            if rag and i % enhance_every == 0:
                arts = rag.retrieve(src)
                ctx = rag.build_context(arts)
            text = f"{ctx}\nسوال: {src}\nپاسخ: {tgt}" if ctx else f"سوال: {src}\nپاسخ: {tgt}"
            self.items.append(text)

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

    def __getitem__(self, idx):
        text = self.items[idx]
        enc = self.tk(text, max_length=self.max_inp, padding="max_length", truncation=True)
        input_ids = torch.tensor(enc["input_ids"])
        attn = torch.tensor(enc["attention_mask"])
        labels = input_ids.clone()
        labels[attn == 0] = -100  # padding mask for loss
        return {"input_ids": input_ids, "attention_mask": attn, "labels": labels}

# ==========================
# Metrics
# ==========================
def build_metrics_fn(arch: str, tokenizer):
    rouge = eval_load("rouge") if eval_load else None

    def _postprocess(preds):
        if isinstance(preds, (list, tuple)):
            return [p.strip() for p in preds]
        return preds

    def compute_metrics_seq2seq(eval_pred):
        if rouge is None:
            return {"rougeL": 0.0}
        preds, labels = eval_pred
        if isinstance(preds, tuple):
            preds = preds[0]
        decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
        labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
        decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
        decoded_preds = _postprocess(decoded_preds)
        decoded_labels = _postprocess(decoded_labels)
        r = rouge.compute(predictions=decoded_preds, references=decoded_labels, rouge_types=["rougeL"])
        return {"rougeL": float(r.get("rougeL", 0.0))}

    def compute_metrics_causal(eval_pred):
        preds, labels = eval_pred
        if isinstance(preds, tuple):
            preds = preds[0]
        decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
        labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
        decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
        tp = fp = fn = 0
        for p, g in zip(decoded_preds, decoded_labels):
            p_set, g_set = set(p.split()), set(g.split())
            tp += len(p_set & g_set)
            fp += len(p_set - g_set)
            fn += len(g_set - p_set)
        precision = tp / (tp + fp + 1e-8)
        recall = tp / (tp + fn + 1e-8)
        f1 = 2 * precision * recall / (precision + recall + 1e-8)
        return {"f1_simple": float(f1)}

    return compute_metrics_seq2seq if arch == "seq2seq" else compute_metrics_causal

# ==========================
# Trainer Manager
# ==========================
def read_jsonl_files(paths: List[str]) -> List[Dict]:
    data: List[Dict] = []
    for p in paths:
        if not p:
            continue
        with open(p, 'r', encoding='utf-8') as f:
            for line in f:
                s = line.strip()
                if not s:
                    continue
                try:
                    obj = json.loads(s)
                    data.append(obj)
                except json.JSONDecodeError:
                    continue
    return data

class TrainerManager:
    def __init__(self, syscfg: SystemConfig, loader: ModelLoader):
        self.cfg = syscfg
        self.loader = loader

    def _args_common(self, is_seq2seq: bool):
        fp16_ok = torch.cuda.is_available() and (not self.cfg.train.use_bf16)
        bf16_ok = bf16_supported() and self.cfg.train.use_bf16

        args = safe_training_args(
            output_dir=self.cfg.train.output_dir,
            num_train_epochs=self.cfg.train.epochs,
            learning_rate=self.cfg.train.lr,
            per_device_train_batch_size=self.cfg.train.batch_size,
            per_device_eval_batch_size=self.cfg.train.batch_size,
            gradient_accumulation_steps=self.cfg.train.grad_accum,
            warmup_ratio=self.cfg.train.warmup_ratio,
            weight_decay=self.cfg.train.weight_decay,
            evaluation_strategy=self.cfg.train.eval_strategy,
            save_strategy=self.cfg.train.save_strategy,
            save_total_limit=self.cfg.train.save_total_limit,
            load_best_model_at_end=True,
            metric_for_best_model="eval_loss",
            logging_steps=self.cfg.train.logging_steps,
            report_to=([] if self.cfg.train.report_to == "none" else [self.cfg.train.report_to]),
            fp16=fp16_ok,
            bf16=bf16_ok,
            max_grad_norm=self.cfg.train.max_grad_norm,
            **({
                "predict_with_generate": True,
                "generation_max_length": self.cfg.model.max_target_length,
                "generation_num_beams": self.cfg.model.num_beams
            } if is_seq2seq else {})
        )
        return args

    def train_seq2seq(self, train_paths: List[str], use_rag: bool = True):
        set_seed_all(self.cfg.train.seed)
        data = read_jsonl_files(train_paths)
        train, val = train_test_split(data, test_size=self.cfg.train.test_size, random_state=self.cfg.train.seed)

        rag = LegalRAG(self.cfg.rag) if (use_rag and self.cfg.rag.enable) else None
        if rag:
            rag.init()

        ds_tr = Seq2SeqJSONLDataset(train, self.loader.tokenizer, self.cfg.model.max_input_length, self.cfg.model.max_target_length, rag)
        ds_va = Seq2SeqJSONLDataset(val,   self.loader.tokenizer, self.cfg.model.max_input_length, self.cfg.model.max_target_length, None)
        collator = DataCollatorForSeq2Seq(tokenizer=self.loader.tokenizer, model=self.loader.model)

        args = self._args_common(is_seq2seq=True)
        trainer = Trainer(
            model=self.loader.model,
            args=args,
            train_dataset=ds_tr,
            eval_dataset=ds_va,
            data_collator=collator,
            tokenizer=self.loader.tokenizer,
            callbacks=[EarlyStoppingCallback(early_stopping_patience=2)],
            compute_metrics=build_metrics_fn("seq2seq", self.loader.tokenizer)
        )
        trainer.train()
        trainer.save_model(self.cfg.train.output_dir)
        self.loader.tokenizer.save_pretrained(self.cfg.train.output_dir)

    def train_causal(self, train_paths: List[str], use_rag: bool = True):
        set_seed_all(self.cfg.train.seed)
        data = read_jsonl_files(train_paths)
        train, val = train_test_split(data, test_size=self.cfg.train.test_size, random_state=self.cfg.train.seed)

        rag = LegalRAG(self.cfg.rag) if (use_rag and self.cfg.rag.enable) else None
        if rag:
            rag.init()

        ds_tr = CausalJSONLDataset(train, self.loader.tokenizer, self.cfg.model.max_input_length, rag)
        ds_va = CausalJSONLDataset(val,   self.loader.tokenizer, self.cfg.model.max_input_length, None)

        args = self._args_common(is_seq2seq=False)
        trainer = Trainer(
            model=self.loader.model,
            args=args,
            train_dataset=ds_tr,
            eval_dataset=ds_va,
            tokenizer=self.loader.tokenizer,
            callbacks=[EarlyStoppingCallback(early_stopping_patience=2)],
            compute_metrics=build_metrics_fn("causal", self.loader.tokenizer)
        )
        trainer.train()
        trainer.save_model(self.cfg.train.output_dir)
        self.loader.tokenizer.save_pretrained(self.cfg.train.output_dir)

# ==========================
# App (Gradio 5)
# ==========================
class LegalApp:
    def __init__(self, scfg: Optional[SystemConfig] = None):
        self.scfg = scfg or SystemConfig()
        self.rag = LegalRAG(self.scfg.rag)
        self.loader: Optional[ModelLoader] = None
        self.gen: Optional[Generator] = None

    # --- helpers ---
    def _file_paths(self, files: List[gr.File]) -> List[str]:
        paths = []
        for f in (files or []):
            p = getattr(f, "name", None) or getattr(f, "path", None)
            if p:
                paths.append(p)
        return paths

    # --- core actions ---
    def load(self, model_name: str, arch: str, use_rag: bool, persist_dir: str, collection: str, top_k: int, threshold: float):
        # configure
        self.scfg.model.model_name = model_name
        self.scfg.model.architecture = arch
        self.scfg.rag.persist_dir = persist_dir
        self.scfg.rag.collection = collection
        self.scfg.rag.top_k = int(top_k)
        self.scfg.rag.similarity_threshold = float(threshold)
        self.scfg.rag.enable = bool(use_rag)

        # load model
        self.loader = ModelLoader(self.scfg.model).load()
        self.gen = Generator(self.loader, self.scfg.model)

        # load rag
        msg_rag = "RAG غیرفعال"
        if use_rag:
            try:
                self.rag = LegalRAG(self.scfg.rag)
                self.rag.init()
                msg_rag = "RAG آماده است"
            except Exception as e:
                msg_rag = f"RAG خطا: {e}"

        return f"مدل بارگذاری شد: {model_name} ({arch})\n{msg_rag}"

    def build_index(self, laws_file: gr.File, id_key: str, text_key: str):
        if not self.scfg.rag.enable:
            return "RAG غیرفعال است."
        try:
            self.rag.init()
            p = getattr(laws_file, "name", None) or getattr(laws_file, "path", None)
            if not p:
                return "فایل قوانین معتبر نیست."
            res = self.rag.index_jsonl(p, id_key=id_key, text_key=text_key)
            return res
        except Exception as e:
            return f"خطا در ایندکس: {e}"

    def answer(self, question: str, use_rag: bool, max_new_tokens: int, temperature: float, top_p: float, num_beams: int):
        if not question.strip():
            return "لطفاً سوال خود را وارد کنید.", ""
        if not self.gen:
            return "ابتدا مدل/RAG را بارگذاری کنید.", ""
        # runtime params
        self.scfg.model.max_new_tokens = int(max_new_tokens)
        self.scfg.model.temperature = float(temperature)
        self.scfg.model.top_p = float(top_p)
        self.scfg.model.num_beams = int(num_beams)
        arts = self.rag.retrieve(question) if (use_rag and self.scfg.rag.enable and self.rag.collection) else []
        ctx = self.rag.build_context(arts) if arts else ""
        ans = self.gen.generate(question, ctx)
        refs = ""
        if arts:
            refs = "\n\n" + "\n".join([f"**ماده {a['article_id']}** (شباهت: {a['similarity']:.2f})\n{a['text'][:380]}..." for a in arts])
        return ans, refs

    def train(self, model_name: str, arch: str, files: List[gr.File], use_rag: bool, epochs: int, batch: int, lr: float,
              wd: float, warmup: float, report_to: str, progress=gr.Progress(track_tqdm=True)):
        progress(0.0, desc="راه‌اندازی")
        self.scfg.model.model_name = model_name
        self.scfg.model.architecture = arch
        self.scfg.train.epochs = int(epochs)
        self.scfg.train.batch_size = int(batch)
        self.scfg.train.lr = float(lr)
        self.scfg.train.weight_decay = float(wd)
        self.scfg.train.warmup_ratio = float(warmup)
        self.scfg.train.report_to = report_to

        progress(0.1, desc="بارگذاری مدل/توکنایزر")
        self.loader = ModelLoader(self.scfg.model).load()

        paths = self._file_paths(files)
        if not paths:
            return "⚠️ هیچ فایل JSONL برای آموزش انتخاب نشده."

        tm = TrainerManager(self.scfg, self.loader)
        set_seed_all(self.scfg.train.seed)

        progress(0.3, desc="آماده‌سازی دیتاست‌ها و RAG")
        if arch == "seq2seq":
            tm.train_seq2seq(paths, use_rag=use_rag)
        else:
            tm.train_causal(paths, use_rag=use_rag)

        progress(0.95, desc="ذخیرهٔ آرتیفکت‌ها")
        return f"✅ آموزش کامل شد و در {self.scfg.train.output_dir} ذخیره شد."

    # --- UI ---
    def build_ui(self):
        log_deps()
        default_models = {
            "Seq2Seq (mt5-base)": ("google/mt5-base", "seq2seq"),
            "Seq2Seq (t5-fa-base)": ("HooshvareLab/t5-fa-base", "seq2seq"),
            "Seq2Seq (flan-t5-base)": ("google/flan-t5-base", "seq2seq"),
            "Causal (Mistral-7B Instruct)": ("mistralai/Mistral-7B-Instruct-v0.2", "causal"),
        }

        with gr.Blocks(title="ماحون — مشاور حقوقی هوشمند", theme=gr.themes.Soft(primary_hue="green", secondary_hue="gray")) as app:
            gr.HTML("""
            <div style='text-align:center;padding:18px'>
              <h1 style='margin-bottom:4px'>ماحون — Ultimate Legal AI</h1>
              <p style='color:#666'>RAG • Seq2Seq/Causal • Training • Metrics</p>
            </div>
            """)

            with gr.Tab("مشاوره"):
                with gr.Row():
                    model_dd = gr.Dropdown(choices=list(default_models.keys()), value="Seq2Seq (mt5-base)", label="مدل")
                    gr.Markdown("**راهنما:** Seq2Seq برای پاسخ‌های ساختاریافته؛ Causal برای مکالمه طبیعی‌تر.")
                with gr.Row():
                    use_rag = gr.Checkbox(value=True, label="RAG فعال باشد؟")
                    persist_dir = gr.Textbox(value=self.scfg.rag.persist_dir, label="مسیر ChromaDB")
                    collection = gr.Textbox(value=self.scfg.rag.collection, label="نام کالکشن")
                with gr.Row():
                    top_k = gr.Slider(1, 15, value=self.scfg.rag.top_k, step=1, label="Top-K")
                    threshold = gr.Slider(0.3, 0.95, value=self.scfg.rag.similarity_threshold, step=0.01, label="آستانه شباهت")
                load_btn = gr.Button("بارگذاری مدل/RAG", variant="primary")
                status = gr.Textbox(label="وضعیت", interactive=False)

                with gr.Accordion("ساخت ایندکس قوانین (اختیاری)", open=False):
                    laws_file = gr.File(label="فایل JSONL قوانین", file_types=[".jsonl"])
                    id_key = gr.Textbox(value="article_id", label="کلید شناسه ماده")
                    text_key = gr.Textbox(value="text", label="کلید متن ماده")
                    index_btn = gr.Button("ایندکس‌سازی قوانین")
                    index_status = gr.Textbox(label="وضعیت ایندکس", interactive=False)

                with gr.Accordion("پارامترهای تولید", open=False):
                    max_new_tokens = gr.Slider(64, 1024, value=self.scfg.model.max_new_tokens, step=16, label="max_new_tokens")
                    temperature = gr.Slider(0.0, 1.5, value=self.scfg.model.temperature, step=0.05, label="temperature")
                    top_p = gr.Slider(0.1, 1.0, value=self.scfg.model.top_p, step=0.05, label="top_p")
                    num_beams = gr.Slider(1, 8, value=self.scfg.model.num_beams, step=1, label="num_beams (Seq2Seq)")

                question = gr.Textbox(lines=3, label="سوال حقوقی")
                gr.Examples(
                    examples=[
                        ["در صورت نقض قرارداد فروش، چه اقداماتی باید انجام دهم؟"],
                        ["آیا درج شرط عدم رقابت در قرارداد کار قانونی است؟"],
                        ["حق و حقوق کارگر در صورت اخراج فوری چیست؟"],
                        ["فرآیند طرح دعوای مطالبه مهریه چگونه است؟"],
                    ],
                    inputs=question, label="نمونه پرسش‌ها"
                )
                ask_btn = gr.Button("پرسش", variant="primary")
                answer = gr.Markdown(label="پاسخ")
                refs = gr.Markdown(label="مواد قانونی مرتبط")

            with gr.Tab("آموزش"):
                gr.Markdown("فایل‌های JSONL با کلیدهای `input` و `output` را بارگذاری کنید.")
                with gr.Row():
                    model_dd_train = gr.Dropdown(choices=list(default_models.keys()), value="Seq2Seq (mt5-base)", label="مدل")
                    use_rag_train = gr.Checkbox(value=True, label="RAG-enhanced Training")
                train_files = gr.Files(label="JSONL Files", file_count="multiple", file_types=[".jsonl"])
                with gr.Row():
                    epochs = gr.Slider(1, 8, value=self.scfg.train.epochs, step=1, label="epochs")
                    batch = gr.Slider(1, 16, value=self.scfg.train.batch_size, step=1, label="batch per device")
                    lr = gr.Number(value=self.scfg.train.lr, label="learning rate")
                with gr.Row():
                    wd = gr.Number(value=self.scfg.train.weight_decay, label="weight decay")
                    warmup = gr.Slider(0.0, 0.2, value=self.scfg.train.warmup_ratio, step=0.01, label="warmup ratio")
                    report_to = gr.Dropdown(choices=["none","wandb"], value=self.scfg.train.report_to, label="report_to")
                train_btn = gr.Button("شروع آموزش", variant="primary")
                train_status = gr.Textbox(label="وضعیت آموزش", interactive=False)

            # رویدادها
            def _resolve(choice: str) -> Tuple[str,str]:
                return default_models[choice]

            load_btn.click(lambda choice, rag, pdir, coll, k, th: self.load(*_resolve(choice), rag, pdir, coll, k, th),
                           inputs=[model_dd, use_rag, persist_dir, collection, top_k, threshold], outputs=status)

            ask_btn.click(lambda q, rag, mnt, t, p, nb: self.answer(q, rag, mnt, t, p, nb),
                          inputs=[question, use_rag, max_new_tokens, temperature, top_p, num_beams],
                          outputs=[answer, refs])

            index_btn.click(lambda f, ik, tk: self.build_index(f, ik, tk),
                            inputs=[laws_file, id_key, text_key], outputs=index_status)

            train_btn.click(
                lambda choice, files, rag, e, b, l, _wd, _wu, _r:
                    self.train(*_resolve(choice), files, rag, e, b, l, _wd, _wu, _r),
                inputs=[model_dd_train, train_files, use_rag_train, epochs, batch, lr, wd, warmup, report_to],
                outputs=train_status
            )
        return app

# ==========================
# Entrypoint for HF Spaces
# ==========================
if __name__ == "__main__":
    app = LegalApp()
    ui = app.build_ui()

    # Gradio 5: بدون concurrency_count
    try:
        ui = ui.queue()  # صف را فعال می‌کند، پارامتر ندارد
    except TypeError:
        # در صورت تفاوت نسخه، ساده لانچ کن
        pass

    ui.launch(server_name="0.0.0.0", server_port=7860)