Upload argos_train_all_datasets.ipynb with huggingface_hub

argos_train_all_datasets.ipynb

@@ -0,0 +1,283 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# ARGOS LoRA Training — AvaSiG/111 Collection\n",
+    "Запускать в Colab с GPU (T4/A100/V100)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# 1. Проверка GPU\n",
+    "!nvidia-smi\n",
+    "import torch\n",
+    "print(f'CUDA: {torch.cuda.is_available()}, устройство: {torch.cuda.get_device_name(0) if torch.cuda.is_available() else \"нет\"}')\n",
+    "print(f'VRAM: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# 2. Установка зависимостей\n",
+    "!pip install -q transformers==4.47.0 peft==0.13.2 trl==0.12.2 bitsandbytes==0.44.1 datasets accelerate huggingface_hub"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": "import os\n\n# HF токен: добавить в Colab через левую панель → секреты → ключ \"HF_TOKEN\"\ntry:\n    from google.colab import userdata\n    HF_TOKEN = userdata.get('HF_TOKEN')\nexcept Exception:\n    HF_TOKEN = os.environ.get('HF_TOKEN', '')\n\nassert HF_TOKEN, \"Добавьте токен HF_TOKEN в Colab Secrets!\"\n\nos.environ['HF_TOKEN'] = HF_TOKEN\nos.environ['HUGGINGFACE_HUB_TOKEN'] = HF_TOKEN\nos.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True'\n\nfrom huggingface_hub import login\nlogin(token=HF_TOKEN, add_to_git_credential=False)\nprint('HF авторизован')"
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# 3. Загрузка и объединение всех датасетов коллекции AvaSiG/111\n",
+    "import torch\n",
+    "from datasets import load_dataset, concatenate_datasets, Dataset\n",
+    "\n",
+    "# Все датасеты из коллекции (без codeparrot — 71GB не помещается, берём 100k сэмпл)\n",
+    "DATASETS = [\n",
+    "    # (repo_id, max_rows или None=всё)\n",
+    "    ('AvaSiG/ru-thinking-reasoning-r1-v2-deduped-bucket', None),\n",
+    "    ('AvaSiG/ru-thinking-reasoning-r1-deduped-bucket',   None),\n",
+    "    ('AvaSiG/ru-tasks-conversation-deduped-bucket',      None),\n",
+    "    ('AvaSiG/ru-instruct-conversation-v3.1-small-bucket', None),\n",
+    "    ('AvaSiG/ru-big-russian-dataset-bucket',             200_000),  # 4GB → сэмпл\n",
+    "    ('AvaSiG/codeparrot-bucket',                         100_000),  # 71GB → сэмпл\n",
+    "]\n",
+    "\n",
+    "SEED = 42\n",
+    "\n",
+    "def extract_text(row, col_names):\n",
+    "    \"\"\"Авто-определение текста по доступным колонкам.\"\"\"\n",
+    "    # Готовый text\n",
+    "    if 'text' in col_names and row.get('text'):\n",
+    "        return row['text']\n",
+    "    # content (big russian dataset)\n",
+    "    if 'content' in col_names and row.get('content'):\n",
+    "        return row['content']\n",
+    "    # chat-формат messages\n",
+    "    if 'messages' in col_names and row.get('messages'):\n",
+    "        msgs = row['messages']\n",
+    "        if isinstance(msgs, str):\n",
+    "            import json\n",
+    "            try: msgs = json.loads(msgs)\n",
+    "            except: return msgs\n",
+    "        txt = ''\n",
+    "        for m in msgs:\n",
+    "            r = m.get('role', '')\n",
+    "            c = m.get('content', '')\n",
+    "            if r == 'system':    txt += f'<|im_start|>system\\n{c}<|im_end|>\\n'\n",
+    "            elif r == 'user':    txt += f'<|im_start|>user\\n{c}<|im_end|>\\n'\n",
+    "            elif r == 'assistant': txt += f'<|im_start|>assistant\\n{c}<|im_end|>\\n'\n",
+    "        return txt\n",
+    "    # conversation колонка\n",
+    "    if 'conversation' in col_names and row.get('conversation'):\n",
+    "        msgs = row['conversation']\n",
+    "        if isinstance(msgs, str):\n",
+    "            import json\n",
+    "            try: msgs = json.loads(msgs)\n",
+    "            except: return msgs\n",
+    "        txt = ''\n",
+    "        for m in msgs:\n",
+    "            r = m.get('role', '') or m.get('from', '')\n",
+    "            c = m.get('content', '') or m.get('value', '')\n",
+    "            if r in ('system', 'gpt'): txt += f'<|im_start|>system\\n{c}<|im_end|>\\n'\n",
+    "            elif r in ('user', 'human'): txt += f'<|im_start|>user\\n{c}<|im_end|>\\n'\n",
+    "            elif r in ('assistant', 'gpt'): txt += f'<|im_start|>assistant\\n{c}<|im_end|>\\n'\n",
+    "        return txt\n",
+    "    # instruction + output\n",
+    "    if 'instruction' in col_names:\n",
+    "        inst = row.get('instruction', '')\n",
+    "        inp  = row.get('input', '')\n",
+    "        out  = row.get('output', '')\n",
+    "        prompt = f'{inst}\\n{inp}'.strip() if inp else inst\n",
+    "        return f'<|im_start|>user\\n{prompt}<|im_end|>\\n<|im_start|>assistant\\n{out}<|im_end|>\\n'\n",
+    "    # code колонка (codeparrot)\n",
+    "    if 'code' in col_names and row.get('code'):\n",
+    "        return row['code']\n",
+    "    return ''\n",
+    "\n",
+    "all_datasets = []\n",
+    "\n",
+    "for repo_id, max_rows in DATASETS:\n",
+    "    print(f'\\nЗагружаю {repo_id}...')\n",
+    "    try:\n",
+    "        ds = load_dataset(repo_id, split='train', token=HF_TOKEN)\n",
+    "        print(f'  строк: {len(ds)}, колонки: {ds.column_names}')\n",
+    "        if max_rows and len(ds) > max_rows:\n",
+    "            ds = ds.shuffle(seed=SEED).select(range(max_rows))\n",
+    "            print(f'  → сэмпл: {len(ds)}')\n",
+    "        col_names = ds.column_names\n",
+    "        ds = ds.map(lambda r: {'text': extract_text(r, col_names)}, remove_columns=col_names)\n",
+    "        ds = ds.filter(lambda x: len(x['text']) > 100)\n",
+    "        print(f'  после фильтра: {len(ds)}')\n",
+    "        all_datasets.append(ds)\n",
+    "    except Exception as e:\n",
+    "        print(f'  ОШИБКА: {e}')\n",
+    "\n",
+    "combined = concatenate_datasets(all_datasets).shuffle(seed=SEED)\n",
+    "split = combined.train_test_split(test_size=0.02, seed=SEED)\n",
+    "train_ds = split['train']\n",
+    "val_ds   = split['test']\n",
+    "print(f'\\nИтого: train={len(train_ds)}, val={len(val_ds)}')\n",
+    "print('Пример:', train_ds[0]['text'][:300])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# 4. Загрузка модели Qwen2.5-7B-Instruct с 4-bit quantization\n",
+    "from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig\n",
+    "from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training\n",
+    "import torch.nn as nn\n",
+    "\n",
+    "if not hasattr(nn.Module, 'set_submodule'):\n",
+    "    def _set_submodule(self, target, module):\n",
+    "        parts = target.split('.')\n",
+    "        parent = self\n",
+    "        for part in parts[:-1]:\n",
+    "            parent = getattr(parent, part)\n",
+    "        setattr(parent, parts[-1], module)\n",
+    "    nn.Module.set_submodule = _set_submodule\n",
+    "\n",
+    "MODEL = 'Qwen/Qwen2.5-7B-Instruct'\n",
+    "OUTPUT = '/content/argos-qwen7b-lora-v3'\n",
+    "\n",
+    "tokenizer = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True, token=HF_TOKEN)\n",
+    "if tokenizer.pad_token is None:\n",
+    "    tokenizer.pad_token = tokenizer.eos_token\n",
+    "\n",
+    "bnb = BitsAndBytesConfig(\n",
+    "    load_in_4bit=True,\n",
+    "    bnb_4bit_quant_type='nf4',\n",
+    "    bnb_4bit_compute_dtype=torch.bfloat16,\n",
+    "    bnb_4bit_use_double_quant=True,\n",
+    ")\n",
+    "model = AutoModelForCausalLM.from_pretrained(\n",
+    "    MODEL,\n",
+    "    quantization_config=bnb,\n",
+    "    device_map={'': 0},\n",
+    "    trust_remote_code=True,\n",
+    "    token=HF_TOKEN,\n",
+    ")\n",
+    "model = prepare_model_for_kbit_training(model)\n",
+    "\n",
+    "lora = LoraConfig(\n",
+    "    r=16,\n",
+    "    lora_alpha=32,\n",
+    "    target_modules=['q_proj','k_proj','v_proj','o_proj','gate_proj','up_proj','down_proj'],\n",
+    "    lora_dropout=0.05,\n",
+    "    bias='none',\n",
+    "    task_type='CAUSAL_LM',\n",
+    ")\n",
+    "model = get_peft_model(model, lora)\n",
+    "model.print_trainable_parameters()\n",
+    "print('Модель готова')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# 5. Обучение\n",
+    "from trl import SFTTrainer, SFTConfig\n",
+    "\n",
+    "args = SFTConfig(\n",
+    "    output_dir=OUTPUT,\n",
+    "    num_train_epochs=3,\n",
+    "    per_device_train_batch_size=1,\n",
+    "    per_device_eval_batch_size=1,\n",
+    "    gradient_accumulation_steps=8,\n",
+    "    learning_rate=2e-4,\n",
+    "    fp16=False,\n",
+    "    bf16=True,\n",
+    "    gradient_checkpointing=True,\n",
+    "    logging_steps=20,\n",
+    "    save_strategy='steps',\n",
+    "    save_steps=500,\n",
+    "    eval_strategy='steps',\n",
+    "    eval_steps=500,\n",
+    "    load_best_model_at_end=False,\n",
+    "    save_total_limit=2,\n",
+    "    remove_unused_columns=False,\n",
+    "    dataloader_num_workers=2,\n",
+    "    lr_scheduler_type='cosine',\n",
+    "    warmup_ratio=0.05,\n",
+    "    dataset_text_field='text',\n",
+    "    max_length=1024,\n",
+    "    pad_to_multiple_of=8,\n",
+    "    push_to_hub=False,\n",
+    "    weight_decay=0.01,\n",
+    "    report_to='none',\n",
+    ")\n",
+    "\n",
+    "trainer = SFTTrainer(\n",
+    "    model=model,\n",
+    "    processing_class=tokenizer,\n",
+    "    train_dataset=train_ds,\n",
+    "    eval_dataset=val_ds,\n",
+    "    args=args,\n",
+    ")\n",
+    "\n",
+    "print(f'Шагов всего: {len(trainer.get_train_dataloader()) * args.num_train_epochs}')\n",
+    "trainer.train()\n",
+    "print('Обучение завершено!')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# 6. Сохранение и загрузка на HF Hub\n",
+    "LORA_OUT = OUTPUT + '/final_lora'\n",
+    "model.save_pretrained(LORA_OUT)\n",
+    "tokenizer.save_pretrained(LORA_OUT)\n",
+    "print(f'Сохранено в {LORA_OUT}')\n",
+    "\n",
+    "from huggingface_hub import HfApi\n",
+    "api = HfApi(token=HF_TOKEN)\n",
+    "api.create_repo('AvaSiG/argos-qwen2.5-7b-v3', exist_ok=True, private=False)\n",
+    "api.upload_folder(folder_path=LORA_OUT, repo_id='AvaSiG/argos-qwen2.5-7b-v3')\n",
+    "print('Готово! Загружено в AvaSiG/argos-qwen2.5-7b-v3')"
+   ]
+  }
+ ],
+ "metadata": {
+  "accelerator": "GPU",
+  "colab": {
+   "gpuType": "T4",
+   "name": "argos_train_all_datasets.ipynb",
+   "provenance": []
+  },
+  "kernelspec": {
+   "display_name": "Python 3",
+   "name": "python3"
+  },
+  "language_info": {
+   "name": "python"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 0
+}