app.py

import os
import gc
import json
import time
import random
import threading
import traceback
from pathlib import Path
from datetime import datetime

import gradio as gr
import psutil
from tqdm.auto import tqdm
from datasets import load_dataset
from tokenizers import Tokenizer
from tokenizers.models import BPE
from tokenizers.trainers import BpeTrainer
from tokenizers.pre_tokenizers import ByteLevel
from tokenizers.decoders import ByteLevel as ByteLevelDecoder
from tokenizers.processors import TemplateProcessing
from huggingface_hub import HfApi, create_repo, upload_folder


# ============================================================
# Env-driven config
# ============================================================

HF_TOKEN = os.environ.get("HF_TOKEN", "").strip()

REPO_ID = os.environ.get("MODEL_REPO_ID", "Efe2898/new-model").strip()
PUSH_TO_HUB = os.environ.get("PUSH_TO_HUB", "1").strip().lower() in {"1", "true", "yes", "y"}

DATASET_ID = os.environ.get("DATASET_ID", "turkish-nlp-suite/BellaTurca").strip()
SUBSETS = [
    s.strip()
    for s in os.environ.get(
        "SUBSETS",
        "AkademikDerlem,OzenliDerlem,temiz-OSCAR,temiz-mC4",
    ).split(",")
    if s.strip()
]
TEXT_COLUMN = os.environ.get("TEXT_COLUMN", "text").strip()
SPLIT = os.environ.get("SPLIT", "train").strip()

VOCAB_SIZE = int(os.environ.get("VOCAB_SIZE", "65536"))
MODEL_MAX_LENGTH = int(os.environ.get("MODEL_MAX_LENGTH", "262144"))
BPE_MIN_FREQUENCY = int(os.environ.get("BPE_MIN_FREQUENCY", "3"))

TARGET_TOTAL_EST_TOKENS = int(os.environ.get("TARGET_TOTAL_EST_TOKENS", "700000000"))
TARGET_PER_SUBSET_EST_TOKENS = TARGET_TOTAL_EST_TOKENS // max(1, len(SUBSETS))
CHARS_PER_EST_TOKEN = float(os.environ.get("CHARS_PER_EST_TOKEN", "4.0"))

CPU_THREADS = int(os.environ.get("CPU_THREADS", "2"))
ROW_BATCH_SIZE = int(os.environ.get("ROW_BATCH_SIZE", "2048"))
MAX_CHARS_PER_BATCH = int(os.environ.get("MAX_CHARS_PER_BATCH", "2000000"))
MAX_TEXT_CHARS = int(os.environ.get("MAX_TEXT_CHARS", "512000"))
MIN_TEXT_CHARS = int(os.environ.get("MIN_TEXT_CHARS", "20"))

ENABLE_CPU_COOLDOWN = os.environ.get("ENABLE_CPU_COOLDOWN", "1").strip().lower() in {"1", "true", "yes", "y"}
MICRO_SLEEP_AFTER_EACH_BATCH = float(os.environ.get("MICRO_SLEEP_AFTER_EACH_BATCH", "0.25"))
COOLDOWN_EVERY_SECONDS = int(os.environ.get("COOLDOWN_EVERY_SECONDS", "60"))
COOLDOWN_SECONDS = float(os.environ.get("COOLDOWN_SECONDS", "5"))
CPU_CHECK_INTERVAL = float(os.environ.get("CPU_CHECK_INTERVAL", "0.10"))
CPU_SOFT_LIMIT_PERCENT = float(os.environ.get("CPU_SOFT_LIMIT_PERCENT", "80"))
CPU_HARD_LIMIT_PERCENT = float(os.environ.get("CPU_HARD_LIMIT_PERCENT", "95"))
CPU_SOFT_SLEEP_SECONDS = float(os.environ.get("CPU_SOFT_SLEEP_SECONDS", "2"))
CPU_HARD_SLEEP_SECONDS = float(os.environ.get("CPU_HARD_SLEEP_SECONDS", "6"))
RAM_HARD_LIMIT_PERCENT = float(os.environ.get("RAM_HARD_LIMIT_PERCENT", "90"))

SHUFFLE_BATCH = os.environ.get("SHUFFLE_BATCH", "0").strip().lower() in {"1", "true", "yes", "y"}
RANDOM_SEED = int(os.environ.get("RANDOM_SEED", "42"))
AUTO_START = os.environ.get("AUTO_START", "0").strip().lower() in {"1", "true", "yes", "y"}

SPACE_PASSWORD = os.environ.get("SPACE_PASSWORD", "738825").strip()

def gradio_auth(username: str, password: str) -> bool:
    """Protect the Space UI. Username is ignored; only the password matters."""
    return str(password).strip() == SPACE_PASSWORD


# Use /data when persistent storage is attached; otherwise Space working dir.
ROOT_DIR = Path(os.environ.get("OUTPUT_ROOT", "/data" if Path("/data").exists() else "."))
OUT_DIR = ROOT_DIR / "rslm_tokenizer_bellaturca_65k"
OUT_DIR.mkdir(parents=True, exist_ok=True)
LOG_PATH = OUT_DIR / "training.log"
STATUS_PATH = OUT_DIR / "status.json"

SPECIAL_TOKENS = [
    "<|pad|>",
    "<|bos|>",
    "<|eos|>",
    "<|unk|>",
    "<|system|>",
    "<|user|>",
    "<|assistant|>",
    "<|answer|>",
    "<|end|>",
    "<think>",
    "</think>",
]

# Keep CPU usage sane.
os.environ["RAYON_NUM_THREADS"] = str(CPU_THREADS)
os.environ["OMP_NUM_THREADS"] = str(CPU_THREADS)
os.environ["MKL_NUM_THREADS"] = str(CPU_THREADS)
os.environ["OPENBLAS_NUM_THREADS"] = str(CPU_THREADS)
os.environ["NUMEXPR_NUM_THREADS"] = str(CPU_THREADS)
os.environ["VECLIB_MAXIMUM_THREADS"] = str(CPU_THREADS)
os.environ["BLIS_NUM_THREADS"] = str(CPU_THREADS)
os.environ["TOKENIZERS_PARALLELISM"] = "false"
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = os.environ.get("HF_HUB_ENABLE_HF_TRANSFER", "0")

_proc = psutil.Process(os.getpid())
psutil.cpu_percent(interval=None)
_proc.cpu_percent(interval=None)
random.seed(RANDOM_SEED)

_state = {
    "thread": None,
    "running": False,
    "finished": False,
    "error": None,
    "uploaded": False,
}


def log(msg: str):
    ts = datetime.utcnow().strftime("%Y-%m-%d %H:%M:%S")
    line = f"[{ts} UTC] {msg}"
    print(line, flush=True)
    with open(LOG_PATH, "a", encoding="utf-8") as f:
        f.write(line + "\n")


def write_status(**kwargs):
    data = {
        "running": _state["running"],
        "finished": _state["finished"],
        "error": _state["error"],
        "uploaded": _state["uploaded"],
        "repo_id": REPO_ID,
        "out_dir": str(OUT_DIR),
        "updated_at_utc": datetime.utcnow().isoformat() + "Z",
    }
    data.update(kwargs)
    data["progress_percent"] = round(calculate_progress_percent(data), 4)

    stats_obj = data.get("stats")
    if isinstance(stats_obj, dict):
        per_subset = {}
        for subset_name, subset_stats in stats_obj.items():
            try:
                subset_est = int(subset_stats.get("est_tokens", 0))
                per_subset[subset_name] = round(
                    max(0.0, min(100.0, 100.0 * subset_est / max(1, TARGET_PER_SUBSET_EST_TOKENS))),
                    4,
                )
            except Exception:
                pass
        data["subset_progress_percent"] = per_subset

    with open(STATUS_PATH, "w", encoding="utf-8") as f:
        json.dump(data, f, ensure_ascii=False, indent=2)



def calculate_progress_percent(status_data=None):
    status_data = status_data or {}
    total = status_data.get("total_est_tokens")
    stats_obj = status_data.get("stats")

    if total is None and isinstance(stats_obj, dict):
        try:
            total = sum(int(v.get("est_tokens", 0)) for v in stats_obj.values())
        except Exception:
            total = 0

    try:
        total = int(total or 0)
    except Exception:
        total = 0

    target = max(1, int(TARGET_TOTAL_EST_TOKENS))
    percent = max(0.0, min(100.0, 100.0 * total / target))
    return percent


def load_status_file():
    if not STATUS_PATH.exists():
        return {}
    try:
        return json.loads(STATUS_PATH.read_text(encoding="utf-8"))
    except Exception:
        return {}

def tail_log(n=220):
    if not LOG_PATH.exists():
        return "Henüz log yok."
    lines = LOG_PATH.read_text(encoding="utf-8", errors="replace").splitlines()
    return "\n".join(lines[-n:])


def system_snapshot():
    vm = psutil.virtual_memory()
    cpu = psutil.cpu_percent(interval=CPU_CHECK_INTERVAL)
    proc_cpu = _proc.cpu_percent(interval=None)
    proc_cpu_norm = proc_cpu / max(1, CPU_THREADS)
    return {
        "cpu": cpu,
        "proc_cpu": proc_cpu,
        "proc_cpu_norm": proc_cpu_norm,
        "ram_percent": vm.percent,
        "rss_gb": _proc.memory_info().rss / (1024 ** 3),
        "available_gb": vm.available / (1024 ** 3),
    }


_last_cooldown_at = time.time()


def maybe_cooldown(reason="batch"):
    global _last_cooldown_at
    if not ENABLE_CPU_COOLDOWN:
        return

    if MICRO_SLEEP_AFTER_EACH_BATCH > 0:
        time.sleep(MICRO_SLEEP_AFTER_EACH_BATCH)

    snap = system_snapshot()
    if snap["ram_percent"] >= RAM_HARD_LIMIT_PERCENT:
        log(
            f"[RAM-FREN] ram={snap['ram_percent']:.1f}% rss={snap['rss_gb']:.2f}GB "
            f"available={snap['available_gb']:.2f}GB -> gc + sleep {CPU_HARD_SLEEP_SECONDS}s"
        )
        gc.collect()
        time.sleep(CPU_HARD_SLEEP_SECONDS)
        _last_cooldown_at = time.time()
        return

    cpu_load = max(snap["cpu"], snap["proc_cpu_norm"])
    if cpu_load >= CPU_HARD_LIMIT_PERCENT:
        log(
            f"[CPU-HARD] reason={reason} cpu={snap['cpu']:.1f}% "
            f"proc_norm={snap['proc_cpu_norm']:.1f}% -> sleep {CPU_HARD_SLEEP_SECONDS}s"
        )
        time.sleep(CPU_HARD_SLEEP_SECONDS)
        _last_cooldown_at = time.time()
    elif cpu_load >= CPU_SOFT_LIMIT_PERCENT:
        log(
            f"[CPU-SOFT] reason={reason} cpu={snap['cpu']:.1f}% "
            f"proc_norm={snap['proc_cpu_norm']:.1f}% -> sleep {CPU_SOFT_SLEEP_SECONDS}s"
        )
        time.sleep(CPU_SOFT_SLEEP_SECONDS)
        _last_cooldown_at = time.time()

    now = time.time()
    if now - _last_cooldown_at >= COOLDOWN_EVERY_SECONDS:
        log(
            f"[COOLDOWN] reason={reason} regular sleep {COOLDOWN_SECONDS}s | "
            f"cpu={snap['cpu']:.1f}% proc_norm={snap['proc_cpu_norm']:.1f}% "
            f"rss={snap['rss_gb']:.2f}GB ram={snap['ram_percent']:.1f}%"
        )
        gc.collect()
        time.sleep(COOLDOWN_SECONDS)
        _last_cooldown_at = time.time()


def clean_text(value):
    if value is None:
        return None
    if not isinstance(value, str):
        value = str(value)
    value = value.strip()
    if len(value) < MIN_TEXT_CHARS:
        return None
    if MAX_TEXT_CHARS and len(value) > MAX_TEXT_CHARS:
        value = value[:MAX_TEXT_CHARS]
    return value


def estimate_tokens(text: str) -> int:
    return max(1, int(len(text) / CHARS_PER_EST_TOKEN))


def train_tokenizer():
    _state["running"] = True
    _state["finished"] = False
    _state["error"] = None
    _state["uploaded"] = False
    write_status(phase="starting")

    stats = {
        subset: {
            "rows": 0,
            "chars": 0,
            "est_tokens": 0,
            "batches": 0,
            "exhausted": False,
            "started_at": None,
            "finished_at": None,
        }
        for subset in SUBSETS
    }

    def total_est_tokens_seen():
        return sum(stats[s]["est_tokens"] for s in SUBSETS)

    def print_stats(prefix=""):
        snap = system_snapshot()
        total_rows = sum(stats[s]["rows"] for s in SUBSETS)
        total_chars = sum(stats[s]["chars"] for s in SUBSETS)
        total_est = total_est_tokens_seen()
        log(
            f"{prefix} cpu={snap['cpu']:.1f}% proc_norm={snap['proc_cpu_norm']:.1f}% "
            f"rss={snap['rss_gb']:.2f}GB ram={snap['ram_percent']:.1f}%"
        )
        log(f"TOTAL rows={total_rows:,} chars={total_chars:,} est_tokens={total_est:,}")
        for subset in SUBSETS:
            st = stats[subset]
            pct = 100 * st["est_tokens"] / max(1, TARGET_PER_SUBSET_EST_TOKENS)
            log(
                f"  {subset:16s} rows={st['rows']:,} batches={st['batches']:,} "
                f"chars={st['chars']:,} est={st['est_tokens']:,} target%={pct:.2f} "
                f"exhausted={st['exhausted']}"
            )
        write_status(phase="training", stats=stats, total_est_tokens=total_est)

    def load_stream(subset: str):
        log(f"Loading stream: {DATASET_ID} / {subset}")
        kwargs = dict(path=DATASET_ID, name=subset, split=SPLIT, streaming=True)
        if HF_TOKEN:
            kwargs["token"] = HF_TOKEN
        return iter(load_dataset(**kwargs))

    def pull_batch(iterator, subset: str):
        batch = []
        batch_chars = 0
        pulled = 0

        while (
            pulled < ROW_BATCH_SIZE
            and batch_chars < MAX_CHARS_PER_BATCH
            and stats[subset]["est_tokens"] < TARGET_PER_SUBSET_EST_TOKENS
            and total_est_tokens_seen() < TARGET_TOTAL_EST_TOKENS
        ):
            try:
                row = next(iterator)
            except StopIteration:
                stats[subset]["exhausted"] = True
                log(f"[WARN] Subset exhausted before target: {subset}")
                break

            text = clean_text(row.get(TEXT_COLUMN))
            if text is None:
                continue

            batch.append(text)
            stats[subset]["rows"] += 1
            stats[subset]["chars"] += len(text)
            stats[subset]["est_tokens"] += estimate_tokens(text)
            batch_chars += len(text)
            pulled += 1

        if batch:
            stats[subset]["batches"] += 1
        return batch

    def text_batches():
        start = time.time()
        last_total = 0
        global_batch_idx = 0

        for subset in SUBSETS:
            if total_est_tokens_seen() >= TARGET_TOTAL_EST_TOKENS:
                break

            stats[subset]["started_at"] = time.time()
            subset_start_est = total_est_tokens_seen()
            stream = load_stream(subset)
            log(f"===== START SUBSET: {subset} =====")

            while (
                stats[subset]["est_tokens"] < TARGET_PER_SUBSET_EST_TOKENS
                and total_est_tokens_seen() < TARGET_TOTAL_EST_TOKENS
                and not stats[subset]["exhausted"]
            ):
                batch = pull_batch(stream, subset)
                if not batch:
                    break

                if SHUFFLE_BATCH:
                    random.shuffle(batch)

                total_est = total_est_tokens_seen()
                global_batch_idx += 1

                if global_batch_idx % 10 == 0:
                    elapsed = max(1e-9, time.time() - start)
                    speed = (total_est - last_total) / max(1e-9, elapsed) / 1e6
                    last_total = total_est
                    print_stats(prefix=f"[batch {global_batch_idx}] elapsed={elapsed/60:.1f}min speed~{speed:.3f}M est_tok/s")

                yield batch
                maybe_cooldown(reason=f"batch_{global_batch_idx}_{subset}")

            stats[subset]["finished_at"] = time.time()
            new_est = total_est_tokens_seen() - subset_start_est
            print_stats(prefix=f"===== FINISH SUBSET: {subset} | new_est={new_est:,} =====")
            del stream
            gc.collect()
            maybe_cooldown(reason=f"finish_{subset}")

        print_stats(prefix="Final iterator stats:")

    try:
        LOG_PATH.parent.mkdir(parents=True, exist_ok=True)
        log("Tokenizer training started.")
        log(f"Repo target: {REPO_ID}")
        log(f"Dataset: {DATASET_ID} / subsets={SUBSETS}")
        log(f"VOCAB_SIZE={VOCAB_SIZE} target_est_tokens={TARGET_TOTAL_EST_TOKENS:,}")
        log(f"CPU_THREADS={CPU_THREADS} ROW_BATCH_SIZE={ROW_BATCH_SIZE} MAX_CHARS_PER_BATCH={MAX_CHARS_PER_BATCH:,}")
        if PUSH_TO_HUB and not HF_TOKEN:
            raise RuntimeError("HF_TOKEN secret/env yok. Push için Space Settings > Secrets içine HF_TOKEN ekle.")

        try:
            bpe_model = BPE(unk_token="<|unk|>", fuse_unk=True, byte_fallback=True)
        except TypeError:
            log("[WARN] tokenizers BPE(byte_fallback=True) desteklemiyor; byte_fallback olmadan devam.")
            bpe_model = BPE(unk_token="<|unk|>", fuse_unk=True)

        tokenizer = Tokenizer(bpe_model)
        tokenizer.pre_tokenizer = ByteLevel(add_prefix_space=False)
        tokenizer.decoder = ByteLevelDecoder()

        trainer = BpeTrainer(
            vocab_size=VOCAB_SIZE,
            min_frequency=BPE_MIN_FREQUENCY,
            special_tokens=SPECIAL_TOKENS,
            initial_alphabet=ByteLevel.alphabet(),
            show_progress=True,
        )

        t0 = time.time()
        tokenizer.train_from_iterator(text_batches(), trainer=trainer, length=TARGET_TOTAL_EST_TOKENS)
        elapsed = time.time() - t0
        log(f"Tokenizer training finished in {elapsed/60:.2f} minutes.")
        log(f"Final vocab size: {tokenizer.get_vocab_size()}")

        bos_id = tokenizer.token_to_id("<|bos|>")
        eos_id = tokenizer.token_to_id("<|eos|>")
        assert bos_id is not None and eos_id is not None, "BOS/EOS token id bulunamadı."

        tokenizer.post_processor = TemplateProcessing(
            single="<|bos|> $A <|eos|>",
            pair="<|bos|> $A <|eos|> $B:1 <|eos|>:1",
            special_tokens=[("<|bos|>", bos_id), ("<|eos|>", eos_id)],
        )

        tokenizer.save(str(OUT_DIR / "tokenizer.json"))

        special_tokens_map = {
            "bos_token": "<|bos|>",
            "eos_token": "<|eos|>",
            "unk_token": "<|unk|>",
            "pad_token": "<|pad|>",
            "additional_special_tokens": [
                "<|system|>",
                "<|user|>",
                "<|assistant|>",
                "<|answer|>",
                "<|end|>",
                "<think>",
                "</think>",
            ],
        }

        chat_template = (
            "{% for message in messages %}"
            "{% if message['role'] == 'system' %}<|system|>\n{{ message['content'] }}<|end|>\n"
            "{% elif message['role'] == 'user' %}<|user|>\n{{ message['content'] }}<|end|>\n"
            "{% elif message['role'] == 'assistant' %}<|assistant|>\n{{ message['content'] }}<|end|>\n"
            "{% endif %}"
            "{% endfor %}"
            "{% if add_generation_prompt %}<|assistant|>\n{% endif %}"
        )

        tokenizer_config = {
            "model_max_length": MODEL_MAX_LENGTH,
            "tokenizer_class": "PreTrainedTokenizerFast",
            "clean_up_tokenization_spaces": False,
            "padding_side": "right",
            "truncation_side": "right",
            "bos_token": "<|bos|>",
            "eos_token": "<|eos|>",
            "unk_token": "<|unk|>",
            "pad_token": "<|pad|>",
            "additional_special_tokens": special_tokens_map["additional_special_tokens"],
            "chat_template": chat_template,
        }

        with open(OUT_DIR / "special_tokens_map.json", "w", encoding="utf-8") as f:
            json.dump(special_tokens_map, f, ensure_ascii=False, indent=2)

        with open(OUT_DIR / "tokenizer_config.json", "w", encoding="utf-8") as f:
            json.dump(tokenizer_config, f, ensure_ascii=False, indent=2)

        report = {
            "created_at_utc": datetime.utcnow().isoformat() + "Z",
            "dataset_id": DATASET_ID,
            "subsets": SUBSETS,
            "split": SPLIT,
            "text_column": TEXT_COLUMN,
            "vocab_size": VOCAB_SIZE,
            "actual_vocab_size": tokenizer.get_vocab_size(),
            "bpe_min_frequency": BPE_MIN_FREQUENCY,
            "model_max_length": MODEL_MAX_LENGTH,
            "target_total_est_tokens": TARGET_TOTAL_EST_TOKENS,
            "target_per_subset_est_tokens": TARGET_PER_SUBSET_EST_TOKENS,
            "chars_per_est_token": CHARS_PER_EST_TOKEN,
            "cpu_threads": CPU_THREADS,
            "row_batch_size": ROW_BATCH_SIZE,
            "max_chars_per_batch": MAX_CHARS_PER_BATCH,
            "max_text_chars": MAX_TEXT_CHARS,
            "stats": stats,
            "special_tokens": SPECIAL_TOKENS,
        }
        with open(OUT_DIR / "tokenizer_training_report.json", "w", encoding="utf-8") as f:
            json.dump(report, f, ensure_ascii=False, indent=2)

        subsets_md = "\n".join(f"- `{s}`" for s in SUBSETS)
        specials_md = "\n".join(f"- `{t}`" for t in SPECIAL_TOKENS)
        readme = f"""---
library_name: tokenizers
language:
- tr
tags:
- turkish
- tokenizer
- byte-level-bpe
- rslm
- think
---

# RSLM Tokenizer 65K

CPU-safe Byte-Level BPE tokenizer for RSLM.

## Training data

Dataset: `{DATASET_ID}`

Subsets:
{subsets_md}

Column: `{TEXT_COLUMN}`

Target estimated tokens: `{TARGET_TOTAL_EST_TOKENS:,}` total, approximately `{TARGET_PER_SUBSET_EST_TOKENS:,}` per subset.

## Vocab

- Requested vocab size: `{VOCAB_SIZE:,}`
- Actual vocab size: `{tokenizer.get_vocab_size():,}`
- BPE min frequency: `{BPE_MIN_FREQUENCY}`

## Special tokens

{specials_md}
"""
        with open(OUT_DIR / "README.md", "w", encoding="utf-8") as f:
            f.write(readme)

        log("Saved files:")
        for p in sorted(OUT_DIR.iterdir()):
            log(f" - {p.name} ({p.stat().st_size / 1024 / 1024:.3f} MB)")

        if PUSH_TO_HUB:
            log("Creating/updating model repo and uploading tokenizer files...")
            create_repo(repo_id=REPO_ID, token=HF_TOKEN, repo_type="model", exist_ok=True)
            upload_folder(
                repo_id=REPO_ID,
                folder_path=str(OUT_DIR),
                token=HF_TOKEN,
                repo_type="model",
                commit_message="Add 65K RSLM tokenizer trained on BellaTurca from Space",
                ignore_patterns=["training.log", "status.json"],
            )
            _state["uploaded"] = True
            log(f"Uploaded tokenizer to: https://huggingface.co/{REPO_ID}")

        _state["finished"] = True
        write_status(phase="finished", stats=stats)
    except Exception as e:
        _state["error"] = repr(e)
        log("ERROR: " + repr(e))
        log(traceback.format_exc())
        write_status(phase="error")
    finally:
        _state["running"] = False


def start_training():
    if _state["running"]:
        return "Zaten çalışıyor.\n\n" + tail_log()
    if _state["finished"]:
        return "Bu oturumda eğitim bitmiş görünüyor. Yeniden başlatmak için Space'i restart edip tekrar deneyebilirsin.\n\n" + tail_log()
    t = threading.Thread(target=train_tokenizer, daemon=True)
    _state["thread"] = t
    t.start()
    return "Eğitim başlatıldı. Logları Refresh ile takip et.\n\n" + tail_log()


def refresh():
    saved_status = load_status_file()
    percent = calculate_progress_percent(saved_status)

    status = {
        "running": _state["running"],
        "finished": _state["finished"],
        "error": _state["error"],
        "uploaded": _state["uploaded"],
        "repo_id": REPO_ID,
        "out_dir": str(OUT_DIR),
        "progress_percent": round(percent, 4),
    }

    if saved_status:
        status.update(saved_status)
        status["running"] = _state["running"]
        status["finished"] = _state["finished"] or saved_status.get("phase") == "finished"
        status["error"] = _state["error"] or saved_status.get("error")
        status["uploaded"] = _state["uploaded"] or bool(saved_status.get("uploaded"))
        percent = calculate_progress_percent(status)
        status["progress_percent"] = round(percent, 4)

    return json.dumps(status, ensure_ascii=False, indent=2), tail_log(), round(percent, 4)


if AUTO_START and not _state["running"] and not _state["finished"]:
    t = threading.Thread(target=train_tokenizer, daemon=True)
    _state["thread"] = t
    t.start()

with gr.Blocks(title="RSLM Tokenizer Trainer") as demo:
    gr.Markdown(
        """
# RSLM Tokenizer Trainer

Bu Space, BellaTurca üzerinden 65K Byte-Level BPE tokenizer eğitir ve sonucu Hugging Face model repo'suna push eder.

Tokeni koda yazma. Space Settings → Secrets içine `HF_TOKEN` olarak ekle.
"""
    )
    with gr.Row():
        start_btn = gr.Button("Start training", variant="primary")
        refresh_btn = gr.Button("Refresh logs")
    progress_box = gr.Slider(
        minimum=0,
        maximum=100,
        value=0,
        step=0.01,
        label="Toplam ilerleme (%)",
        interactive=False,
    )
    status_box = gr.Code(label="Status", language="json")
    log_box = gr.Textbox(label="Logs", lines=28, max_lines=50)

    start_btn.click(start_training, outputs=log_box)
    refresh_btn.click(refresh, outputs=[status_box, log_box, progress_box])
    demo.load(refresh, outputs=[status_box, log_box, progress_box])

demo.queue(default_concurrency_limit=1).launch(auth=gradio_auth)