Upload 05_generation_finetune.ipynb

05_generation_finetune.ipynb

@@ -0,0 +1,124 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Mise — Part 5: Generation + LoRA fine-tune on real recipes (bonus +10%)\n",
+    "\n",
+    "The **generation** component + the fine-tuning bonus. We fine-tune `Qwen2.5-3B-Instruct` (QLoRA) on our **10,000 real recipes** so it becomes a **cuisine-specialized recipe generator**:\n",
+    "\n",
+    "> instruction: *\"Write a {difficulty} {cuisine} {dish_type} recipe.\"*  →  the recipe\n",
+    "\n",
+    "This is genuine domain fine-tuning on real data. The app uses this model to generate the '1 new recipe' for a cuisine/level.\n",
+    "\n",
+    "**Needs a GPU (Colab T4 / Kaggle).** Reads the dataset from HF; pushes the model to `idoyaaran/mise-lesson-model`."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 1. Load dataset from HF"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": "!pip -q install -U transformers accelerate peft bitsandbytes datasets huggingface_hub\nimport json, torch, pandas as pd\nfrom huggingface_hub import hf_hub_download\n\n# download the parquet directly (robust — avoids the fragile hf:// fsspec protocol)\npath = hf_hub_download(\"idoyaaran/mise-recipes\", \"recipes_clean.parquet\", repo_type=\"dataset\")\ndf = pd.read_parquet(path)\nBASE_MODEL = \"Qwen/Qwen2.5-3B-Instruct\"\nprint(len(df), \"recipes\")"
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 2. Format recipes into (instruction → recipe) training pairs\n",
+    "\n",
+    "The instruction is what the app will send; the output is the real recipe written as clean text."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def instruction(r):\n",
+    "    return f\"Write a {r['difficulty']} {r['cuisine'].replace('_',' ')} {r['dish_type'].replace('_',' ')} recipe.\"\n",
+    "\n",
+    "def format_recipe(r):\n",
+    "    ings = \"\\n\".join(f\"- {i['name']} ({i.get('quantity','')})\" for i in r['ingredients'])\n",
+    "    steps = \"\\n\".join(f\"{n}. {s}\" for n, s in enumerate(r['steps'], 1))\n",
+    "    return (f\"Title: {r['title']}\\n\"\n",
+    "            f\"Cuisine: {r['cuisine']} | Difficulty: {r['difficulty']} | Time: {r['time_minutes']} min | Serves: {r['servings']}\\n\"\n",
+    "            f\"Ingredients:\\n{ings}\\n\"\n",
+    "            f\"Steps:\\n{steps}\")\n",
+    "\n",
+    "pairs = [{\"instruction\": instruction(r), \"output\": format_recipe(r)} for r in df.to_dict(\"records\")]\n",
+    "train = pd.DataFrame(pairs)\n",
+    "print(len(train), \"training pairs\\n\")\n",
+    "print(\"INSTRUCTION:\", train.iloc[0][\"instruction\"])\n",
+    "print(\"\\nOUTPUT:\\n\", train.iloc[0][\"output\"][:400])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 3. Load base model (4-bit / QLoRA) + baseline generation"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": "from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig\nbnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.float16, bnb_4bit_quant_type=\"nf4\")\ntok = AutoTokenizer.from_pretrained(BASE_MODEL)\nmodel = AutoModelForCausalLM.from_pretrained(BASE_MODEL, quantization_config=bnb, device_map=\"auto\")\n\ndef generate(instr, max_new_tokens=400):\n    msgs = [{\"role\": \"user\", \"content\": instr}]\n    inp = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors=\"pt\", return_dict=True).to(model.device)\n    out = model.generate(**inp, max_new_tokens=max_new_tokens, do_sample=True, temperature=0.7,\n                         top_p=0.9, repetition_penalty=1.15, use_cache=True,\n                         pad_token_id=tok.eos_token_id)\n    return tok.decode(out[0][inp[\"input_ids\"].shape[1]:], skip_special_tokens=True)\n\nprint(\"BEFORE fine-tuning:\\n\")\nprint(generate(\"Write a beginner mexican street food recipe.\"))"
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": "## 4. LoRA fine-tune (transformers Trainer + peft — no trl)"
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": "from datasets import Dataset\nfrom peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training\nfrom transformers import TrainingArguments, Trainer, DataCollatorForSeq2Seq\n\nif tok.pad_token is None:\n    tok.pad_token = tok.eos_token\n\n# subsample for faster + more controllable training (plenty to specialize the format)\ntrain_sub = train.sample(n=min(3000, len(train)), random_state=42).reset_index(drop=True)\n\nmodel = prepare_model_for_kbit_training(model)\npeft_cfg = LoraConfig(r=16, lora_alpha=32, lora_dropout=0.05, task_type=\"CAUSAL_LM\",\n                      target_modules=[\"q_proj\",\"k_proj\",\"v_proj\",\"o_proj\"])\nmodel = get_peft_model(model, peft_cfg)\nmodel.enable_input_require_grads()\nmodel.print_trainable_parameters()\n\nMAXLEN = 640\ndef tokenize(ex):\n    prompt_text = tok.apply_chat_template(\n        [{\"role\": \"user\", \"content\": ex[\"instruction\"]}],\n        add_generation_prompt=True, tokenize=False)\n    full_text = tok.apply_chat_template(\n        [{\"role\": \"user\", \"content\": ex[\"instruction\"]},\n         {\"role\": \"assistant\", \"content\": ex[\"output\"]}], tokenize=False)\n    prompt_ids = tok(prompt_text, add_special_tokens=False)[\"input_ids\"]\n    full_ids = tok(full_text, add_special_tokens=False, truncation=True, max_length=MAXLEN)[\"input_ids\"]\n    labels = full_ids.copy()\n    for i in range(min(len(prompt_ids), len(labels))):\n        labels[i] = -100\n    return {\"input_ids\": full_ids, \"attention_mask\": [1] * len(full_ids), \"labels\": labels}\n\nds = Dataset.from_pandas(train_sub).map(tokenize, remove_columns=list(train_sub.columns))\n\nargs = TrainingArguments(\n    output_dir=\"mise-lesson-model\",\n    per_device_train_batch_size=4,\n    gradient_accumulation_steps=4,\n    num_train_epochs=1,\n    learning_rate=1e-4,\n    warmup_ratio=0.05,\n    lr_scheduler_type=\"cosine\",\n    max_grad_norm=0.3,                 # aggressive clipping -> prevents divergence\n    optim=\"paged_adamw_8bit\",          # stable QLoRA optimizer\n    logging_steps=20,\n    fp16=True,\n    gradient_checkpointing=True,\n    gradient_checkpointing_kwargs={\"use_reentrant\": False},\n    report_to=\"none\",\n    save_strategy=\"no\",\n)\nmodel.config.use_cache = False\ntrainer = Trainer(model=model, args=args, train_dataset=ds,\n                  data_collator=DataCollatorForSeq2Seq(tok, padding=True))\ntrainer.train()\nmodel.save_pretrained(\"mise-lesson-model\")\ntok.save_pretrained(\"mise-lesson-model\")\nprint(\"Saved LoRA adapter.\")"
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 5. Test the fine-tuned model, then push to HF"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": "model.eval()\nmodel.config.use_cache = True\nprint(\"AFTER fine-tuning:\\n\")\nprint(generate(\"Write an advanced italian main recipe.\"))"
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": "from huggingface_hub import login\ntry:\n    from kaggle_secrets import UserSecretsClient      # Kaggle\n    login(UserSecretsClient().get_secret(\"HF_TOKEN\"))\nexcept Exception:\n    try:\n        from google.colab import userdata              # Colab\n        login(userdata.get(\"HF_TOKEN\"))\n    except Exception:\n        login()                                        # interactive prompt\n\nMODEL_REPO = \"idoyaaran/mise-lesson-model\"\ntrainer.model.push_to_hub(MODEL_REPO)\ntok.push_to_hub(MODEL_REPO)\nprint(\"Pushed ->\", f\"https://huggingface.co/{MODEL_REPO}\")"
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "name": "python3"
+  },
+  "language_info": {
+   "name": "python"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 0
+}