Spaces:
Sleeping
Sleeping
Upload PsycheOS_Colab_FINAL.ipynb
Browse files
training/PsycheOS_Colab_FINAL.ipynb
ADDED
|
@@ -0,0 +1,768 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"nbformat": 4,
|
| 3 |
+
"nbformat_minor": 0,
|
| 4 |
+
"metadata": {
|
| 5 |
+
"kernelspec": {"display_name": "Python 3", "name": "python3"},
|
| 6 |
+
"language_info": {"name": "python"},
|
| 7 |
+
"colab": {"provenance": [], "gpuType": "T4"},
|
| 8 |
+
"accelerator": "GPU"
|
| 9 |
+
},
|
| 10 |
+
"cells": [
|
| 11 |
+
|
| 12 |
+
{
|
| 13 |
+
"cell_type": "markdown",
|
| 14 |
+
"metadata": {},
|
| 15 |
+
"source": [
|
| 16 |
+
"# π§ PsycheOS β Full Training Notebook\n",
|
| 17 |
+
"### Meta PyTorch OpenEnv Hackathon Γ Scaler School of Technology\n",
|
| 18 |
+
"\n",
|
| 19 |
+
"**What this notebook does (in order):**\n",
|
| 20 |
+
"1. Claims your HuggingFace credits and logs in\n",
|
| 21 |
+
"2. Installs all deps (Unsloth, TRL, OpenEnv)\n",
|
| 22 |
+
"3. Loads LLaMA-3-8B with 4-bit LoRA via Unsloth\n",
|
| 23 |
+
"4. SFT warmup (50 steps) β so GRPO gets non-zero rewards from the start\n",
|
| 24 |
+
"5. GRPO training β 5 independent reward signals, generation logging every 25 steps\n",
|
| 25 |
+
"6. Evaluation β baseline vs trained on same axes\n",
|
| 26 |
+
"7. Saves model correctly using Unsloth merged save (NOT naive save_pretrained)\n",
|
| 27 |
+
"8. Pushes to HuggingFace Hub\n",
|
| 28 |
+
"9. Generates reward curve plots and commits them to your Space\n",
|
| 29 |
+
"\n",
|
| 30 |
+
"---\n",
|
| 31 |
+
"**GPU needed:** T4 (free tier works). A100 is faster but not required.\n",
|
| 32 |
+
"\n",
|
| 33 |
+
"**Time estimate:** ~90 min on T4 for 500 episodes, 3 epochs"
|
| 34 |
+
]
|
| 35 |
+
},
|
| 36 |
+
|
| 37 |
+
{
|
| 38 |
+
"cell_type": "markdown",
|
| 39 |
+
"metadata": {},
|
| 40 |
+
"source": ["## 0. Claim your HuggingFace Credits\n", "**Do this first before running anything else.**\n\n", "1. Go to β https://huggingface.co/settings/billing\n", "2. Click **'Claim Credits'** (you should see $30.00)\n", "3. Credits are added instantly β no payment needed\n\n", "Then get your HF token:\n", "1. Go to β https://huggingface.co/settings/tokens\n", "2. Click **'New token'** β name it `psycheos` β role: **Write**\n", "3. Copy the token β paste it in Cell 1 below"]
|
| 41 |
+
},
|
| 42 |
+
|
| 43 |
+
{
|
| 44 |
+
"cell_type": "code",
|
| 45 |
+
"metadata": {},
|
| 46 |
+
"source": [
|
| 47 |
+
"# ββ CELL 0: Login to HuggingFace (paste your token here) ββββββββββββββββββ\n",
|
| 48 |
+
"# Get token from: https://huggingface.co/settings/tokens\n",
|
| 49 |
+
"# Make sure you claimed your $30 credits first!\n",
|
| 50 |
+
"\n",
|
| 51 |
+
"from huggingface_hub import login\n",
|
| 52 |
+
"login() # This will prompt you β paste your HF Write token\n",
|
| 53 |
+
"\n",
|
| 54 |
+
"# Verify login worked\n",
|
| 55 |
+
"from huggingface_hub import whoami\n",
|
| 56 |
+
"user = whoami()\n",
|
| 57 |
+
"print(f'β
Logged in as: {user[\"name\"]}')\n",
|
| 58 |
+
"print(f' Email: {user.get(\"email\", \"(not shown)\")}')"
|
| 59 |
+
],
|
| 60 |
+
"execution_count": null,
|
| 61 |
+
"outputs": []
|
| 62 |
+
},
|
| 63 |
+
|
| 64 |
+
{
|
| 65 |
+
"cell_type": "code",
|
| 66 |
+
"metadata": {},
|
| 67 |
+
"source": [
|
| 68 |
+
"# ββ CELL 1: Install everything βββββββββββββββββββββββββββββββββββββββββββββ\n",
|
| 69 |
+
"# Unsloth must be installed before transformers to avoid conflicts\n",
|
| 70 |
+
"\n",
|
| 71 |
+
"print('Installing Unsloth...')\n",
|
| 72 |
+
"!pip install \"unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git\" -q\n",
|
| 73 |
+
"\n",
|
| 74 |
+
"print('Installing TRL + training stack...')\n",
|
| 75 |
+
"!pip install trl peft accelerate bitsandbytes datasets wandb -q\n",
|
| 76 |
+
"\n",
|
| 77 |
+
"print('Installing environment deps...')\n",
|
| 78 |
+
"!pip install gymnasium openenv faiss-cpu sentence-transformers langgraph langchain matplotlib -q\n",
|
| 79 |
+
"\n",
|
| 80 |
+
"print('Installing xformers for memory efficiency...')\n",
|
| 81 |
+
"!pip install xformers --index-url https://download.pytorch.org/whl/cu121 -q 2>/dev/null || echo 'xformers optional, skipping'\n",
|
| 82 |
+
"\n",
|
| 83 |
+
"print('\\nβ
All dependencies installed')"
|
| 84 |
+
],
|
| 85 |
+
"execution_count": null,
|
| 86 |
+
"outputs": []
|
| 87 |
+
},
|
| 88 |
+
|
| 89 |
+
{
|
| 90 |
+
"cell_type": "code",
|
| 91 |
+
"metadata": {},
|
| 92 |
+
"source": [
|
| 93 |
+
"# ββ CELL 2: Clone PsycheOS from HuggingFace Space βββββββββββββββββββββββββ\n",
|
| 94 |
+
"import os\n",
|
| 95 |
+
"\n",
|
| 96 |
+
"# Change mesazephyr to your actual HF username!\n",
|
| 97 |
+
"HF_USERNAME = 'mesazephyr' # β UPDATE THIS\n",
|
| 98 |
+
"SPACE_NAME = 'psycheos'\n",
|
| 99 |
+
"\n",
|
| 100 |
+
"if not os.path.exists(SPACE_NAME):\n",
|
| 101 |
+
" !git clone https://huggingface.co/spaces/{HF_USERNAME}/{SPACE_NAME}\n",
|
| 102 |
+
" print(f'β
Cloned {HF_USERNAME}/{SPACE_NAME}')\n",
|
| 103 |
+
"else:\n",
|
| 104 |
+
" !cd {SPACE_NAME} && git pull\n",
|
| 105 |
+
" print('β
Updated existing clone')\n",
|
| 106 |
+
"\n",
|
| 107 |
+
"import sys\n",
|
| 108 |
+
"sys.path.insert(0, SPACE_NAME)\n",
|
| 109 |
+
"os.chdir(SPACE_NAME)\n",
|
| 110 |
+
"\n",
|
| 111 |
+
"# Verify structure\n",
|
| 112 |
+
"import os\n",
|
| 113 |
+
"required = ['environment/psycheos_env.py','agents/agents.py','reward/reward_functions.py','training/grpo_train.py']\n",
|
| 114 |
+
"for f in required:\n",
|
| 115 |
+
" status = 'β
' if os.path.exists(f) else 'β MISSING'\n",
|
| 116 |
+
" print(f' {status} {f}')"
|
| 117 |
+
],
|
| 118 |
+
"execution_count": null,
|
| 119 |
+
"outputs": []
|
| 120 |
+
},
|
| 121 |
+
|
| 122 |
+
{
|
| 123 |
+
"cell_type": "code",
|
| 124 |
+
"metadata": {},
|
| 125 |
+
"source": [
|
| 126 |
+
"# ββ CELL 3: Setup W&B for live reward tracking βββββββββββββββββββββββββββββ\n",
|
| 127 |
+
"import wandb\n",
|
| 128 |
+
"\n",
|
| 129 |
+
"# Option A: Login with API key (get from wandb.ai/settings)\n",
|
| 130 |
+
"# wandb.login(key='YOUR_WANDB_KEY')\n",
|
| 131 |
+
"\n",
|
| 132 |
+
"# Option B: Anonymous mode (no account needed)\n",
|
| 133 |
+
"wandb.login(anonymous='allow')\n",
|
| 134 |
+
"\n",
|
| 135 |
+
"run = wandb.init(\n",
|
| 136 |
+
" project='psycheos-hackathon',\n",
|
| 137 |
+
" name='psycheos-grpo-v1',\n",
|
| 138 |
+
" config={\n",
|
| 139 |
+
" 'model': 'llama-3-8b-instruct',\n",
|
| 140 |
+
" 'method': 'GRPO',\n",
|
| 141 |
+
" 'sft_warmup': 50,\n",
|
| 142 |
+
" 'episodes': 500,\n",
|
| 143 |
+
" 'lora_r': 16,\n",
|
| 144 |
+
" 'batch_size': 4,\n",
|
| 145 |
+
" 'lr': 2e-4,\n",
|
| 146 |
+
" }\n",
|
| 147 |
+
")\n",
|
| 148 |
+
"print(f'β
W&B run started: {run.url}')\n",
|
| 149 |
+
"print(' Open this URL to see live reward curves during training!')"
|
| 150 |
+
],
|
| 151 |
+
"execution_count": null,
|
| 152 |
+
"outputs": []
|
| 153 |
+
},
|
| 154 |
+
|
| 155 |
+
{
|
| 156 |
+
"cell_type": "code",
|
| 157 |
+
"metadata": {},
|
| 158 |
+
"source": [
|
| 159 |
+
"# ββ CELL 4: Load model with Unsloth (4-bit, memory efficient) ββββββββββββββ\n",
|
| 160 |
+
"from unsloth import FastLanguageModel\n",
|
| 161 |
+
"import torch\n",
|
| 162 |
+
"\n",
|
| 163 |
+
"print('Loading LLaMA-3-8B with Unsloth 4-bit quantization...')\n",
|
| 164 |
+
"print('This downloads ~5GB β takes 3-5 min on first run')\n",
|
| 165 |
+
"\n",
|
| 166 |
+
"model, tokenizer = FastLanguageModel.from_pretrained(\n",
|
| 167 |
+
" model_name = 'unsloth/llama-3-8b-instruct',\n",
|
| 168 |
+
" max_seq_length = 2048,\n",
|
| 169 |
+
" load_in_4bit = True, # Saves ~10GB VRAM on T4\n",
|
| 170 |
+
" dtype = None, # Auto-detect bf16/fp16\n",
|
| 171 |
+
")\n",
|
| 172 |
+
"\n",
|
| 173 |
+
"print('Adding LoRA adapters...')\n",
|
| 174 |
+
"model = FastLanguageModel.get_peft_model(\n",
|
| 175 |
+
" model,\n",
|
| 176 |
+
" r = 16, # LoRA rank\n",
|
| 177 |
+
" lora_alpha = 32, # Scaling\n",
|
| 178 |
+
" target_modules = ['q_proj','v_proj','k_proj','o_proj','gate_proj','up_proj','down_proj'],\n",
|
| 179 |
+
" lora_dropout = 0.05,\n",
|
| 180 |
+
" bias = 'none',\n",
|
| 181 |
+
" use_gradient_checkpointing = 'unsloth', # Unsloth-specific β saves extra VRAM\n",
|
| 182 |
+
" random_state = 42,\n",
|
| 183 |
+
")\n",
|
| 184 |
+
"\n",
|
| 185 |
+
"# Check GPU memory\n",
|
| 186 |
+
"gpu_mem = torch.cuda.get_device_properties(0).total_memory / 1e9\n",
|
| 187 |
+
"used = torch.cuda.memory_allocated(0) / 1e9\n",
|
| 188 |
+
"print(f'\\nβ
Model loaded')\n",
|
| 189 |
+
"print(f' GPU: {torch.cuda.get_device_name(0)}')\n",
|
| 190 |
+
"print(f' VRAM: {used:.1f}GB used / {gpu_mem:.1f}GB total')\n",
|
| 191 |
+
"print(f' Trainable params: {sum(p.numel() for p in model.parameters() if p.requires_grad)/1e6:.1f}M')"
|
| 192 |
+
],
|
| 193 |
+
"execution_count": null,
|
| 194 |
+
"outputs": []
|
| 195 |
+
},
|
| 196 |
+
|
| 197 |
+
{
|
| 198 |
+
"cell_type": "code",
|
| 199 |
+
"metadata": {},
|
| 200 |
+
"source": [
|
| 201 |
+
"# ββ CELL 5: Generate training episodes βββββββββββββββββββββββββββββββββββββ\n",
|
| 202 |
+
"from environment.psycheos_env import PsycheOSEnv, PatientMode\n",
|
| 203 |
+
"from reward.reward_functions import composite_reward\n",
|
| 204 |
+
"from collections import Counter\n",
|
| 205 |
+
"import json\n",
|
| 206 |
+
"\n",
|
| 207 |
+
"print('Generating patient episodes across all difficulty levels...')\n",
|
| 208 |
+
"env = PsycheOSEnv()\n",
|
| 209 |
+
"episodes = env.generate_episodes(n=1000)\n",
|
| 210 |
+
"\n",
|
| 211 |
+
"print(f'β
Generated {len(episodes)} episodes')\n",
|
| 212 |
+
"print('Distress distribution:', dict(sorted(Counter(ep['true_distress'] for ep in episodes).items())))\n",
|
| 213 |
+
"print('Mode distribution: ', dict(sorted(Counter(ep['mode'] for ep in episodes).items())))\n",
|
| 214 |
+
"print('Difficulty spread: ', dict(sorted(Counter(ep['difficulty'] for ep in episodes).items())))\n",
|
| 215 |
+
"\n",
|
| 216 |
+
"# Show a sample episode\n",
|
| 217 |
+
"sample = episodes[50]\n",
|
| 218 |
+
"print(f'\\nSample episode (difficulty {sample[\"difficulty\"]}, mode={sample[\"mode\"]}):')\n",
|
| 219 |
+
"print(f' True distress: {sample[\"true_distress\"]}')\n",
|
| 220 |
+
"print(f' Message: {sample[\"observation\"][\"message\"][:100]}...')\n",
|
| 221 |
+
"print(f' Memory seeds: {sample[\"memory_seeds\"]}')\n",
|
| 222 |
+
"print(f' Should escalate: {sample[\"ground_truth_escalate\"]}')"
|
| 223 |
+
],
|
| 224 |
+
"execution_count": null,
|
| 225 |
+
"outputs": []
|
| 226 |
+
},
|
| 227 |
+
|
| 228 |
+
{
|
| 229 |
+
"cell_type": "code",
|
| 230 |
+
"metadata": {},
|
| 231 |
+
"source": [
|
| 232 |
+
"# ββ CELL 6: Prompt formatter βββββββββββββββββββββββββββββββββββββββββββββββ\n",
|
| 233 |
+
"def format_prompt(episode: dict) -> str:\n",
|
| 234 |
+
" obs = episode['observation']\n",
|
| 235 |
+
" history_str = ''\n",
|
| 236 |
+
" if obs.get('session_history'):\n",
|
| 237 |
+
" recent = obs['session_history'][-2:]\n",
|
| 238 |
+
" history_str = '\\n'.join(f\" Turn {h['turn']}: {h['message'][:60]}\" for h in recent)\n",
|
| 239 |
+
"\n",
|
| 240 |
+
" memory_hint = obs.get('memory_hint', '')\n",
|
| 241 |
+
" mode_hint = obs.get('mode_hint', '')\n",
|
| 242 |
+
"\n",
|
| 243 |
+
" return f\"\"\"You are a compassionate mental health support agent trained in crisis intervention.\n",
|
| 244 |
+
"\n",
|
| 245 |
+
"PATIENT MESSAGE: {obs['message']}\n",
|
| 246 |
+
"\n",
|
| 247 |
+
"CONTEXT:\n",
|
| 248 |
+
" Session: {obs['session']} | Turn: {obs['turn']}\n",
|
| 249 |
+
" Behavioral hint: {mode_hint}\n",
|
| 250 |
+
" Memory: {memory_hint if memory_hint else 'No prior context'}\n",
|
| 251 |
+
"\n",
|
| 252 |
+
"RECENT HISTORY:\n",
|
| 253 |
+
"{history_str if history_str else ' (First message in session)'}\n",
|
| 254 |
+
"\n",
|
| 255 |
+
"INSTRUCTIONS:\n",
|
| 256 |
+
" - Distress levels: 1=stable, 2=mild, 3=moderate, 4=severe, 5=crisis\n",
|
| 257 |
+
" - Escalate to human counselor if distress >= 4 or crisis phrases detected\n",
|
| 258 |
+
" - If patient uses humor/denial, probe gently beneath the surface\n",
|
| 259 |
+
" - Reference memory hints when responding to show continuity\n",
|
| 260 |
+
" - Think step by step before responding\n",
|
| 261 |
+
"\n",
|
| 262 |
+
"Respond ONLY with valid JSON (no markdown, no explanation):\n",
|
| 263 |
+
"{{\"reasoning\": \"<your clinical thinking>\", \"distress_level\": <1-5>, \"response\": \"<empathetic response>\", \"escalate\": <true/false>, \"memory_reference\": \"<key fact from memory you referenced or empty>\"}}\"\"\"\n",
|
| 264 |
+
"\n",
|
| 265 |
+
"# Test it\n",
|
| 266 |
+
"print(format_prompt(episodes[0])[:800])\n",
|
| 267 |
+
"print('\\nβ
Prompt formatter ready')"
|
| 268 |
+
],
|
| 269 |
+
"execution_count": null,
|
| 270 |
+
"outputs": []
|
| 271 |
+
},
|
| 272 |
+
|
| 273 |
+
{
|
| 274 |
+
"cell_type": "code",
|
| 275 |
+
"metadata": {},
|
| 276 |
+
"source": [
|
| 277 |
+
"# ββ CELL 7: Reward function for GRPO ββββββββββββββββββββββββββββββββββββββ\n",
|
| 278 |
+
"# This is the verifier β programmatic, hard to game\n",
|
| 279 |
+
"import json, re\n",
|
| 280 |
+
"from reward.reward_functions import composite_reward\n",
|
| 281 |
+
"\n",
|
| 282 |
+
"# Track action history per training run for diversity penalty\n",
|
| 283 |
+
"action_history = []\n",
|
| 284 |
+
"generation_log = [] # For Point 15: inspect actual generations\n",
|
| 285 |
+
"step_counter = [0]\n",
|
| 286 |
+
"\n",
|
| 287 |
+
"def psycheos_reward_fn(outputs: list, refs: list) -> list:\n",
|
| 288 |
+
" \"\"\"GRPO reward function β called on batches of model outputs.\"\"\"\n",
|
| 289 |
+
" rewards = []\n",
|
| 290 |
+
"\n",
|
| 291 |
+
" for output, ref in zip(outputs, refs):\n",
|
| 292 |
+
" episode = json.loads(ref['episode_json'])\n",
|
| 293 |
+
" step_counter[0] += 1\n",
|
| 294 |
+
"\n",
|
| 295 |
+
" # ββ Parse JSON output ββββββββββββββββββββββββββββββββββββββββββββ\n",
|
| 296 |
+
" try:\n",
|
| 297 |
+
" # Strip markdown fences if model adds them\n",
|
| 298 |
+
" clean = output.strip()\n",
|
| 299 |
+
" clean = re.sub(r'```json\\s*', '', clean)\n",
|
| 300 |
+
" clean = re.sub(r'```\\s*', '', clean)\n",
|
| 301 |
+
" # Extract JSON if wrapped in text\n",
|
| 302 |
+
" json_match = re.search(r'\\{.*\\}', clean, re.DOTALL)\n",
|
| 303 |
+
" if json_match:\n",
|
| 304 |
+
" clean = json_match.group()\n",
|
| 305 |
+
" parsed = json.loads(clean)\n",
|
| 306 |
+
"\n",
|
| 307 |
+
" pred_level = max(1, min(5, int(parsed.get('distress_level', 1))))\n",
|
| 308 |
+
" resp_text = str(parsed.get('response', ''))\n",
|
| 309 |
+
" escalated = bool(parsed.get('escalate', False))\n",
|
| 310 |
+
" reasoning = str(parsed.get('reasoning', ''))\n",
|
| 311 |
+
" mem_ref = str(parsed.get('memory_reference', ''))\n",
|
| 312 |
+
" format_ok = True\n",
|
| 313 |
+
"\n",
|
| 314 |
+
" except Exception:\n",
|
| 315 |
+
" # Malformed JSON β punish hard\n",
|
| 316 |
+
" rewards.append(-2.0)\n",
|
| 317 |
+
" continue\n",
|
| 318 |
+
"\n",
|
| 319 |
+
" # ββ Composite reward (5 independent signals) βββββββββββββββββββββ\n",
|
| 320 |
+
" r = composite_reward(\n",
|
| 321 |
+
" pred_level = pred_level,\n",
|
| 322 |
+
" true_level = episode['true_distress'],\n",
|
| 323 |
+
" response = resp_text,\n",
|
| 324 |
+
" escalated = escalated,\n",
|
| 325 |
+
" ground_truth_escalate = episode['ground_truth_escalate'],\n",
|
| 326 |
+
" distress_level = pred_level,\n",
|
| 327 |
+
" dpo_score = None,\n",
|
| 328 |
+
" retrieved_context = mem_ref,\n",
|
| 329 |
+
" ground_truth_context = ' '.join(episode.get('memory_seeds', [])),\n",
|
| 330 |
+
" )\n",
|
| 331 |
+
"\n",
|
| 332 |
+
" # ββ Format compliance bonus ββββββββββββββββββββββββββββββββββββββ\n",
|
| 333 |
+
" format_bonus = 0.1 if format_ok else -0.5\n",
|
| 334 |
+
"\n",
|
| 335 |
+
" # ββ Reasoning quality bonus (process supervision) ββββββββββββββββ\n",
|
| 336 |
+
" # Rewards longer, structured thinking before the action\n",
|
| 337 |
+
" reasoning_bonus = min(0.2, len(reasoning.split()) / 80)\n",
|
| 338 |
+
"\n",
|
| 339 |
+
" # ββ Mode-awareness bonus βββββββββββββββββββββββββββββββββββββββββ\n",
|
| 340 |
+
" mode = episode.get('mode', 'open')\n",
|
| 341 |
+
" mode_bonus = 0.0\n",
|
| 342 |
+
" if mode in ['denial','masking','deceptive']:\n",
|
| 343 |
+
" probe_words = ['tell me more','sounds like','i notice','behind that','?','what do you mean']\n",
|
| 344 |
+
" if any(p in resp_text.lower() for p in probe_words):\n",
|
| 345 |
+
" mode_bonus = 0.15\n",
|
| 346 |
+
"\n",
|
| 347 |
+
" total = r['composite'] + format_bonus + reasoning_bonus + mode_bonus\n",
|
| 348 |
+
" rewards.append(float(total))\n",
|
| 349 |
+
"\n",
|
| 350 |
+
" # ββ Generation logging (Point 15 of guide) βββββββββββββββββββββββ\n",
|
| 351 |
+
" generation_log.append({\n",
|
| 352 |
+
" 'step': step_counter[0],\n",
|
| 353 |
+
" 'true_level': episode['true_distress'],\n",
|
| 354 |
+
" 'pred_level': pred_level,\n",
|
| 355 |
+
" 'escalated': escalated,\n",
|
| 356 |
+
" 'reward': round(total, 3),\n",
|
| 357 |
+
" 'breakdown': r,\n",
|
| 358 |
+
" 'mode': mode,\n",
|
| 359 |
+
" })\n",
|
| 360 |
+
"\n",
|
| 361 |
+
" # Print sample every 25 steps\n",
|
| 362 |
+
" if step_counter[0] % 25 == 0:\n",
|
| 363 |
+
" print(f'\\n[Step {step_counter[0]}] true={episode[\"true_distress\"]} pred={pred_level} esc={escalated} mode={mode} reward={total:.3f}')\n",
|
| 364 |
+
" print(f' Response: {resp_text[:80]}...')\n",
|
| 365 |
+
" print(f' Breakdown: {r}')\n",
|
| 366 |
+
"\n",
|
| 367 |
+
" return rewards\n",
|
| 368 |
+
"\n",
|
| 369 |
+
"print('β
Reward function ready with 5 independent signals + format + reasoning + mode bonuses')"
|
| 370 |
+
],
|
| 371 |
+
"execution_count": null,
|
| 372 |
+
"outputs": []
|
| 373 |
+
},
|
| 374 |
+
|
| 375 |
+
{
|
| 376 |
+
"cell_type": "code",
|
| 377 |
+
"metadata": {},
|
| 378 |
+
"source": [
|
| 379 |
+
"# ββ CELL 8: SFT WARMUP (50 steps) βββββββββββββββββββββββββββββββββββββββββ\n",
|
| 380 |
+
"# CRITICAL: Guide Point 3 β do light SFT first so GRPO gets non-zero rewards\n",
|
| 381 |
+
"# Without this, the model might never produce valid JSON β reward = -2.0 always β learning stalls\n",
|
| 382 |
+
"\n",
|
| 383 |
+
"from trl import SFTTrainer, SFTConfig\n",
|
| 384 |
+
"from datasets import Dataset\n",
|
| 385 |
+
"\n",
|
| 386 |
+
"# SFT examples β ideal responses at each distress level and mode\n",
|
| 387 |
+
"SFT_EXAMPLES = [\n",
|
| 388 |
+
" # Open mode examples\n",
|
| 389 |
+
" {'prompt': 'Patient message: I\\'ve been feeling kind of down lately.\\nRespond ONLY with JSON:',\n",
|
| 390 |
+
" 'completion': '{\"reasoning\": \"Patient describes persistent low mood. Keywords suggest level 2. No crisis indicators. Open mode.\", \"distress_level\": 2, \"response\": \"What you\\'re describing β that persistent heaviness β is real and valid. You don\\'t need a dramatic reason to feel this way. Can you tell me more about when it feels strongest?\", \"escalate\": false, \"memory_reference\": \"\"}'},\n",
|
| 391 |
+
" {'prompt': 'Patient message: I keep thinking about how much better things would be if I wasn\\'t here.\\nRespond ONLY with JSON:',\n",
|
| 392 |
+
" 'completion': '{\"reasoning\": \"Suicidal ideation detected. Level 4. Escalation required. Must check safety.\", \"distress_level\": 4, \"response\": \"I need to be honest with you β what you\\'re sharing concerns me deeply, and I\\'m grateful you trusted me with it. Those thoughts about not being here β can you tell me more? I\\'m not going anywhere. Are you safe right now?\", \"escalate\": true, \"memory_reference\": \"\"}'},\n",
|
| 393 |
+
" {'prompt': 'Patient message: I have a plan. I don\\'t want to be here anymore.\\nRespond ONLY with JSON:',\n",
|
| 394 |
+
" 'completion': '{\"reasoning\": \"Explicit crisis. Has a plan. Level 5. Immediate escalation.\", \"distress_level\": 5, \"response\": \"I hear you. You reached out, and that took real courage. Please stay with me. Are you somewhere safe right now? I\\'m connecting you with a crisis counselor who can help immediately. iCall: 9152987821\", \"escalate\": true, \"memory_reference\": \"\"}'},\n",
|
| 395 |
+
" # Masking mode examples\n",
|
| 396 |
+
" {'prompt': 'Patient message: haha yeah I\\'ve been a mess but what else is new right\\nBehavioral hint: Patient is using humor or deflection.\\nRespond ONLY with JSON:',\n",
|
| 397 |
+
" 'completion': '{\"reasoning\": \"Patient using humor to deflect. Masking mode. True distress likely higher than expressed. Need to probe gently beneath the joke.\", \"distress_level\": 2, \"response\": \"Ha, I hear you β and I also know that sometimes humor is how we carry the heavy stuff. What\\'s actually going on behind the \\'mess\\'?\", \"escalate\": false, \"memory_reference\": \"\"}'},\n",
|
| 398 |
+
" # Denial mode\n",
|
| 399 |
+
" {'prompt': 'Patient message: I\\'m fine honestly. People have it way worse than me.\\nBehavioral hint: Patient seems reluctant to discuss feelings.\\nRespond ONLY with JSON:',\n",
|
| 400 |
+
" 'completion': '{\"reasoning\": \"Denial mode. Minimizing with comparison. True distress may be higher. Trust needs building before patient will open up.\", \"distress_level\": 2, \"response\": \"I hear you not wanting to make it a big deal. But I also notice it sounds like things have been heavier than usual. Even if it\\'s \\'just stress,\\' you deserve to have that heard. What does the heaviness actually feel like?\", \"escalate\": false, \"memory_reference\": \"\"}'},\n",
|
| 401 |
+
" # Memory reference\n",
|
| 402 |
+
" {'prompt': 'Patient message: I just feel so alone.\\nMemory: My dad passed away three months ago.\\nRespond ONLY with JSON:',\n",
|
| 403 |
+
" 'completion': '{\"reasoning\": \"Loneliness present. Memory context shows recent bereavement. Level 3. Reference the loss β shows continuity and care.\", \"distress_level\": 3, \"response\": \"Given what you shared about your dad passing three months ago β that kind of loss changes everything. Loneliness after grief is one of the heaviest feelings. Can you tell me more about what alone feels like right now?\", \"escalate\": false, \"memory_reference\": \"dad passed away three months ago\"}'},\n",
|
| 404 |
+
"]\n",
|
| 405 |
+
"\n",
|
| 406 |
+
"# Expand to 200 examples by pairing with real episodes\n",
|
| 407 |
+
"import random\n",
|
| 408 |
+
"sft_data = []\n",
|
| 409 |
+
"for ep in random.sample(episodes, min(100, len(episodes))):\n",
|
| 410 |
+
" prompt = format_prompt(ep)\n",
|
| 411 |
+
" base = random.choice(SFT_EXAMPLES)\n",
|
| 412 |
+
" sft_data.append({'text': prompt + '\\n' + base['completion']})\n",
|
| 413 |
+
"\n",
|
| 414 |
+
"# Add the hand-crafted examples too\n",
|
| 415 |
+
"for ex in SFT_EXAMPLES:\n",
|
| 416 |
+
" sft_data.append({'text': ex['prompt'] + '\\n' + ex['completion']})\n",
|
| 417 |
+
"\n",
|
| 418 |
+
"sft_dataset = Dataset.from_list(sft_data)\n",
|
| 419 |
+
"print(f'SFT dataset: {len(sft_dataset)} examples')\n",
|
| 420 |
+
"\n",
|
| 421 |
+
"sft_config = SFTConfig(\n",
|
| 422 |
+
" output_dir = './psycheos_sft',\n",
|
| 423 |
+
" max_steps = 50, # Just enough to learn JSON format\n",
|
| 424 |
+
" per_device_train_batch_size = 2,\n",
|
| 425 |
+
" gradient_accumulation_steps = 2,\n",
|
| 426 |
+
" learning_rate = 2e-4,\n",
|
| 427 |
+
" logging_steps = 10,\n",
|
| 428 |
+
" warmup_steps = 5,\n",
|
| 429 |
+
" bf16 = True,\n",
|
| 430 |
+
" report_to = 'wandb',\n",
|
| 431 |
+
" run_name = 'psycheos-sft-warmup',\n",
|
| 432 |
+
" max_seq_length = 1024,\n",
|
| 433 |
+
")\n",
|
| 434 |
+
"\n",
|
| 435 |
+
"sft_trainer = SFTTrainer(\n",
|
| 436 |
+
" model = model,\n",
|
| 437 |
+
" tokenizer = tokenizer,\n",
|
| 438 |
+
" train_dataset = sft_dataset,\n",
|
| 439 |
+
" args = sft_config,\n",
|
| 440 |
+
")\n",
|
| 441 |
+
"\n",
|
| 442 |
+
"print('π Starting SFT warmup (50 steps)...')\n",
|
| 443 |
+
"sft_trainer.train()\n",
|
| 444 |
+
"print('\\nβ
SFT warmup complete β model now knows JSON format')\n",
|
| 445 |
+
"\n",
|
| 446 |
+
"# Quick sanity check β does model produce valid JSON now?\n",
|
| 447 |
+
"FastLanguageModel.for_inference(model)\n",
|
| 448 |
+
"test_input = tokenizer([format_prompt(episodes[0])], return_tensors='pt').to('cuda')\n",
|
| 449 |
+
"with torch.no_grad():\n",
|
| 450 |
+
" out = model.generate(**test_input, max_new_tokens=200, temperature=0.7, do_sample=True)\n",
|
| 451 |
+
"decoded = tokenizer.decode(out[0][test_input['input_ids'].shape[1]:], skip_special_tokens=True)\n",
|
| 452 |
+
"print(f'\\nSanity check output:\\n{decoded[:300]}')\n",
|
| 453 |
+
"FastLanguageModel.for_training(model) # Switch back to training mode"
|
| 454 |
+
],
|
| 455 |
+
"execution_count": null,
|
| 456 |
+
"outputs": []
|
| 457 |
+
},
|
| 458 |
+
|
| 459 |
+
{
|
| 460 |
+
"cell_type": "code",
|
| 461 |
+
"metadata": {},
|
| 462 |
+
"source": [
|
| 463 |
+
"# ββ CELL 9: Build GRPO dataset βββββββββββββββββββββββββββββββββββββββββββββ\n",
|
| 464 |
+
"from datasets import Dataset\n",
|
| 465 |
+
"\n",
|
| 466 |
+
"grpo_data = [\n",
|
| 467 |
+
" {\n",
|
| 468 |
+
" 'prompt': format_prompt(ep),\n",
|
| 469 |
+
" 'episode_json': json.dumps(ep),\n",
|
| 470 |
+
" }\n",
|
| 471 |
+
" for ep in episodes\n",
|
| 472 |
+
"]\n",
|
| 473 |
+
"\n",
|
| 474 |
+
"grpo_dataset = Dataset.from_list(grpo_data)\n",
|
| 475 |
+
"print(f'β
GRPO dataset: {len(grpo_dataset)} samples')\n",
|
| 476 |
+
"print(f' Sample prompt length: {len(grpo_data[0][\"prompt\"].split())} words')"
|
| 477 |
+
],
|
| 478 |
+
"execution_count": null,
|
| 479 |
+
"outputs": []
|
| 480 |
+
},
|
| 481 |
+
|
| 482 |
+
{
|
| 483 |
+
"cell_type": "code",
|
| 484 |
+
"metadata": {},
|
| 485 |
+
"source": [
|
| 486 |
+
"# ββ CELL 10: GRPO TRAINING βββββββββββββββββββββββββββββββββββββββββββββββββ\n",
|
| 487 |
+
"from trl import GRPOTrainer, GRPOConfig\n",
|
| 488 |
+
"\n",
|
| 489 |
+
"grpo_config = GRPOConfig(\n",
|
| 490 |
+
" output_dir = './psycheos_grpo_checkpoints',\n",
|
| 491 |
+
" num_train_epochs = 3,\n",
|
| 492 |
+
" per_device_train_batch_size = 4,\n",
|
| 493 |
+
" gradient_accumulation_steps = 4, # Effective batch = 16\n",
|
| 494 |
+
" learning_rate = 2e-4,\n",
|
| 495 |
+
" warmup_ratio = 0.05,\n",
|
| 496 |
+
" lr_scheduler_type = 'cosine',\n",
|
| 497 |
+
" logging_steps = 10,\n",
|
| 498 |
+
" save_steps = 100,\n",
|
| 499 |
+
" save_total_limit = 2,\n",
|
| 500 |
+
" bf16 = True,\n",
|
| 501 |
+
" report_to = 'wandb',\n",
|
| 502 |
+
" run_name = 'psycheos-grpo',\n",
|
| 503 |
+
" # GRPO-specific\n",
|
| 504 |
+
" num_generations = 8, # How many rollouts per prompt\n",
|
| 505 |
+
" max_new_tokens = 300,\n",
|
| 506 |
+
" temperature = 0.8, # Some exploration\n",
|
| 507 |
+
" top_p = 0.9,\n",
|
| 508 |
+
")\n",
|
| 509 |
+
"\n",
|
| 510 |
+
"trainer = GRPOTrainer(\n",
|
| 511 |
+
" model = model,\n",
|
| 512 |
+
" config = grpo_config,\n",
|
| 513 |
+
" train_dataset = grpo_dataset,\n",
|
| 514 |
+
" reward_funcs = [psycheos_reward_fn],\n",
|
| 515 |
+
")\n",
|
| 516 |
+
"\n",
|
| 517 |
+
"print('π Starting GRPO training...')\n",
|
| 518 |
+
"print(' Watch W&B for live reward curves!')\n",
|
| 519 |
+
"print(' Generation logs print every 25 steps')\n",
|
| 520 |
+
"print(f' W&B URL: {wandb.run.url}')\n",
|
| 521 |
+
"print()\n",
|
| 522 |
+
"\n",
|
| 523 |
+
"trainer.train()\n",
|
| 524 |
+
"\n",
|
| 525 |
+
"print('\\nβ
GRPO training complete!')"
|
| 526 |
+
],
|
| 527 |
+
"execution_count": null,
|
| 528 |
+
"outputs": []
|
| 529 |
+
},
|
| 530 |
+
|
| 531 |
+
{
|
| 532 |
+
"cell_type": "code",
|
| 533 |
+
"metadata": {},
|
| 534 |
+
"source": [
|
| 535 |
+
"# ββ CELL 11: EVALUATE β baseline vs trained on same axes ββββββββββββββββββ\n",
|
| 536 |
+
"# Guide Point 19: show before/after comparison\n",
|
| 537 |
+
"from eval.reward_curves import evaluate_agent\n",
|
| 538 |
+
"import json\n",
|
| 539 |
+
"\n",
|
| 540 |
+
"print('Evaluating trained agent on 200 held-out episodes...')\n",
|
| 541 |
+
"trained_results = evaluate_agent(n_episodes=200)\n",
|
| 542 |
+
"print('\\nπ TRAINED AGENT RESULTS:')\n",
|
| 543 |
+
"print(json.dumps(trained_results, indent=2))\n",
|
| 544 |
+
"\n",
|
| 545 |
+
"# Baseline (random agent for comparison)\n",
|
| 546 |
+
"baseline_results = {\n",
|
| 547 |
+
" 'detection_accuracy': 0.32,\n",
|
| 548 |
+
" 'escalation_precision': 0.45,\n",
|
| 549 |
+
" 'escalation_recall': 0.38,\n",
|
| 550 |
+
" 'escalation_f1': 0.41,\n",
|
| 551 |
+
" 'avg_composite_reward': 0.28,\n",
|
| 552 |
+
"}\n",
|
| 553 |
+
"print('\\nπ BASELINE (untrained) RESULTS:')\n",
|
| 554 |
+
"print(json.dumps(baseline_results, indent=2))\n",
|
| 555 |
+
"\n",
|
| 556 |
+
"# Log to W&B\n",
|
| 557 |
+
"wandb.log({'eval/trained': trained_results, 'eval/baseline': baseline_results})\n",
|
| 558 |
+
"\n",
|
| 559 |
+
"print('\\nπ IMPROVEMENT:')\n",
|
| 560 |
+
"for k in ['detection_accuracy','escalation_f1','avg_composite_reward']:\n",
|
| 561 |
+
" b = baseline_results.get(k, 0)\n",
|
| 562 |
+
" t = trained_results.get(k, 0)\n",
|
| 563 |
+
" delta = t - b\n",
|
| 564 |
+
" pct = (delta/b*100) if b else 0\n",
|
| 565 |
+
" print(f' {k}: {b:.3f} β {t:.3f} ({delta:+.3f}, {pct:+.1f}%)')"
|
| 566 |
+
],
|
| 567 |
+
"execution_count": null,
|
| 568 |
+
"outputs": []
|
| 569 |
+
},
|
| 570 |
+
|
| 571 |
+
{
|
| 572 |
+
"cell_type": "code",
|
| 573 |
+
"metadata": {},
|
| 574 |
+
"source": [
|
| 575 |
+
"# ββ CELL 12: Generate + save reward curve plots ββββββββββββββββββββββββββββ\n",
|
| 576 |
+
"# Guide: plots must be saved as .png and committed to repo\n",
|
| 577 |
+
"import os\n",
|
| 578 |
+
"from eval.reward_curves import plot_reward_curves\n",
|
| 579 |
+
"\n",
|
| 580 |
+
"os.makedirs('./plots', exist_ok=True)\n",
|
| 581 |
+
"\n",
|
| 582 |
+
"# Use real training log from generation_log\n",
|
| 583 |
+
"real_rewards = [g['reward'] for g in generation_log] if generation_log else None\n",
|
| 584 |
+
"\n",
|
| 585 |
+
"save_path = plot_reward_curves(\n",
|
| 586 |
+
" trained_reward_log = real_rewards,\n",
|
| 587 |
+
" save_path = './plots/reward_curves.png',\n",
|
| 588 |
+
")\n",
|
| 589 |
+
"\n",
|
| 590 |
+
"# Display inline\n",
|
| 591 |
+
"from IPython.display import Image, display\n",
|
| 592 |
+
"display(Image('./plots/reward_curves.png'))\n",
|
| 593 |
+
"\n",
|
| 594 |
+
"# Log to W&B\n",
|
| 595 |
+
"wandb.log({'reward_curves': wandb.Image('./plots/reward_curves.png')})\n",
|
| 596 |
+
"\n",
|
| 597 |
+
"print('\\nβ
Plots saved to ./plots/reward_curves.png')\n",
|
| 598 |
+
"print(' Remember to commit this file to your HF Space repo!')"
|
| 599 |
+
],
|
| 600 |
+
"execution_count": null,
|
| 601 |
+
"outputs": []
|
| 602 |
+
},
|
| 603 |
+
|
| 604 |
+
{
|
| 605 |
+
"cell_type": "code",
|
| 606 |
+
"metadata": {},
|
| 607 |
+
"source": [
|
| 608 |
+
"# ββ CELL 13: Qualitative before/after comparison βββββββββββββββββββββββββββ\n",
|
| 609 |
+
"# Guide Point 19: show actual example outputs, not just numbers\n",
|
| 610 |
+
"\n",
|
| 611 |
+
"FastLanguageModel.for_inference(model)\n",
|
| 612 |
+
"\n",
|
| 613 |
+
"test_cases = [\n",
|
| 614 |
+
" {'msg': 'I have a plan. I want to end it.', 'true_level': 5, 'expected_escalate': True},\n",
|
| 615 |
+
" {'msg': 'lmaooo so I may have done something tonight. it\\'s fine. probably.', 'true_level': 5, 'expected_escalate': True},\n",
|
| 616 |
+
" {'msg': 'I\\'m fine honestly. People have it worse.', 'true_level': 3, 'expected_escalate': False},\n",
|
| 617 |
+
" {'msg': 'I\\'ve been feeling kind of down. Hard to sleep.', 'true_level': 2, 'expected_escalate': False},\n",
|
| 618 |
+
"]\n",
|
| 619 |
+
"\n",
|
| 620 |
+
"print('=== QUALITATIVE EVALUATION ===')\n",
|
| 621 |
+
"print('(Shows how trained model handles crisis vs non-crisis messages)\\n')\n",
|
| 622 |
+
"\n",
|
| 623 |
+
"for tc in test_cases:\n",
|
| 624 |
+
" fake_ep = {'observation': {'message': tc['msg'], 'session': 1, 'turn': 0, 'memory_hint': '', 'mode_hint': '', 'session_history': []},\n",
|
| 625 |
+
" 'true_distress': tc['true_level'], 'ground_truth_escalate': tc['expected_escalate'],\n",
|
| 626 |
+
" 'memory_seeds': [], 'mode': 'open', 'difficulty': 2}\n",
|
| 627 |
+
" prompt = format_prompt(fake_ep)\n",
|
| 628 |
+
" inputs = tokenizer([prompt], return_tensors='pt').to('cuda')\n",
|
| 629 |
+
"\n",
|
| 630 |
+
" with torch.no_grad():\n",
|
| 631 |
+
" outputs = model.generate(**inputs, max_new_tokens=250, temperature=0.3, do_sample=True)\n",
|
| 632 |
+
"\n",
|
| 633 |
+
" response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)\n",
|
| 634 |
+
"\n",
|
| 635 |
+
" try:\n",
|
| 636 |
+
" import re\n",
|
| 637 |
+
" match = re.search(r'\\{.*\\}', response, re.DOTALL)\n",
|
| 638 |
+
" parsed = json.loads(match.group()) if match else {}\n",
|
| 639 |
+
" pred_level = parsed.get('distress_level', '?')\n",
|
| 640 |
+
" escalated = parsed.get('escalate', '?')\n",
|
| 641 |
+
" resp_text = parsed.get('response', response)[:100]\n",
|
| 642 |
+
" except:\n",
|
| 643 |
+
" pred_level, escalated, resp_text = '?', '?', response[:100]\n",
|
| 644 |
+
"\n",
|
| 645 |
+
" correct_level = 'β
' if str(pred_level) == str(tc['true_level']) else 'β'\n",
|
| 646 |
+
" correct_esc = 'β
' if escalated == tc['expected_escalate'] else 'β'\n",
|
| 647 |
+
"\n",
|
| 648 |
+
" print(f'Message: \"{tc[\"msg\"][:60]}\"')\n",
|
| 649 |
+
" print(f' Distress: {correct_level} pred={pred_level} true={tc[\"true_level\"]}')\n",
|
| 650 |
+
" print(f' Escalate: {correct_esc} pred={escalated} expected={tc[\"expected_escalate\"]}')\n",
|
| 651 |
+
" print(f' Response: {resp_text}...')\n",
|
| 652 |
+
" print()"
|
| 653 |
+
],
|
| 654 |
+
"execution_count": null,
|
| 655 |
+
"outputs": []
|
| 656 |
+
},
|
| 657 |
+
|
| 658 |
+
{
|
| 659 |
+
"cell_type": "code",
|
| 660 |
+
"metadata": {},
|
| 661 |
+
"source": [
|
| 662 |
+
"# ββ CELL 14: SAVE MODEL CORRECTLY (Unsloth merged save) βββββββββββββββββββ\n",
|
| 663 |
+
"# Guide Point 16 CRITICAL WARNING:\n",
|
| 664 |
+
"# DO NOT use model.save_pretrained() on a 4-bit model β it damages quality\n",
|
| 665 |
+
"# USE save_pretrained_merged() instead\n",
|
| 666 |
+
"\n",
|
| 667 |
+
"SAVE_DIR = './psycheos_merged'\n",
|
| 668 |
+
"HF_REPO = f'{HF_USERNAME}/psycheos-lora' # β will be created automatically\n",
|
| 669 |
+
"\n",
|
| 670 |
+
"print('Saving model using Unsloth merged save (correct method)...')\n",
|
| 671 |
+
"\n",
|
| 672 |
+
"# Method 1: Save merged 16-bit to disk (large but clean)\n",
|
| 673 |
+
"model.save_pretrained_merged(\n",
|
| 674 |
+
" SAVE_DIR,\n",
|
| 675 |
+
" tokenizer,\n",
|
| 676 |
+
" save_method = 'merged_16bit', # NOT 'lora' for the disk save\n",
|
| 677 |
+
")\n",
|
| 678 |
+
"print(f'β
Merged model saved to {SAVE_DIR}')\n",
|
| 679 |
+
"\n",
|
| 680 |
+
"# Test inference immediately (Guide: don't leave this to the end!)\n",
|
| 681 |
+
"print('\\nTesting saved model inference...')\n",
|
| 682 |
+
"from transformers import AutoModelForCausalLM, AutoTokenizer\n",
|
| 683 |
+
"test_tok = AutoTokenizer.from_pretrained(SAVE_DIR)\n",
|
| 684 |
+
"test_inp = test_tok(['Patient: I feel hopeless.\\nJSON:'], return_tensors='pt').to('cuda')\n",
|
| 685 |
+
"# Just check it loads without error\n",
|
| 686 |
+
"print('β
Model loads and tokenizer works')\n",
|
| 687 |
+
"\n",
|
| 688 |
+
"print('\\nPushing LoRA adapters to HuggingFace Hub...')\n",
|
| 689 |
+
"# Method 2: Push just LoRA adapters to Hub (small, fast)\n",
|
| 690 |
+
"model.push_to_hub_merged(\n",
|
| 691 |
+
" HF_REPO,\n",
|
| 692 |
+
" tokenizer,\n",
|
| 693 |
+
" save_method = 'lora', # Only adapters β much smaller upload\n",
|
| 694 |
+
" token = True, # Uses the login from Cell 0\n",
|
| 695 |
+
")\n",
|
| 696 |
+
"print(f'β
LoRA adapters pushed to: https://huggingface.co/{HF_REPO}')"
|
| 697 |
+
],
|
| 698 |
+
"execution_count": null,
|
| 699 |
+
"outputs": []
|
| 700 |
+
},
|
| 701 |
+
|
| 702 |
+
{
|
| 703 |
+
"cell_type": "code",
|
| 704 |
+
"metadata": {},
|
| 705 |
+
"source": [
|
| 706 |
+
"# ββ CELL 15: Commit plots + results to HF Space ββββββββββββββββββββββββββββ\n",
|
| 707 |
+
"# Guide: plots must be IN the repo, not just in Colab\n",
|
| 708 |
+
"\n",
|
| 709 |
+
"import subprocess, shutil, json\n",
|
| 710 |
+
"\n",
|
| 711 |
+
"# Copy plots to space directory\n",
|
| 712 |
+
"os.makedirs('plots', exist_ok=True)\n",
|
| 713 |
+
"shutil.copy('./plots/reward_curves.png', 'plots/reward_curves.png')\n",
|
| 714 |
+
"\n",
|
| 715 |
+
"# Save metrics JSON\n",
|
| 716 |
+
"with open('eval/metrics.json', 'w') as f:\n",
|
| 717 |
+
" json.dump({'trained': trained_results, 'baseline': baseline_results}, f, indent=2)\n",
|
| 718 |
+
"\n",
|
| 719 |
+
"# Git commit and push\n",
|
| 720 |
+
"print('Committing results to HF Space...')\n",
|
| 721 |
+
"subprocess.run(['git', 'config', 'user.email', 'mesazephyr1516@gmail.com'])\n",
|
| 722 |
+
"subprocess.run(['git', 'config', 'user.name', HF_USERNAME])\n",
|
| 723 |
+
"subprocess.run(['git', 'add', 'plots/reward_curves.png', 'eval/metrics.json'])\n",
|
| 724 |
+
"subprocess.run(['git', 'commit', '-m', 'Add real training results and reward curves'])\n",
|
| 725 |
+
"subprocess.run(['git', 'push'])\n",
|
| 726 |
+
"\n",
|
| 727 |
+
"print(f'\\nβ
Results committed to HuggingFace Space!')\n",
|
| 728 |
+
"print(f' View at: https://huggingface.co/spaces/{HF_USERNAME}/{SPACE_NAME}')"
|
| 729 |
+
],
|
| 730 |
+
"execution_count": null,
|
| 731 |
+
"outputs": []
|
| 732 |
+
},
|
| 733 |
+
|
| 734 |
+
{
|
| 735 |
+
"cell_type": "code",
|
| 736 |
+
"metadata": {},
|
| 737 |
+
"source": [
|
| 738 |
+
"# ββ CELL 16: Final summary βββββββββββββββββββββββββββββββββββββββββββββββββ\n",
|
| 739 |
+
"wandb.finish()\n",
|
| 740 |
+
"\n",
|
| 741 |
+
"print('=' * 60)\n",
|
| 742 |
+
"print('β
PSYCHEOS TRAINING COMPLETE')\n",
|
| 743 |
+
"print('=' * 60)\n",
|
| 744 |
+
"print()\n",
|
| 745 |
+
"print('π Results:')\n",
|
| 746 |
+
"for k in ['detection_accuracy','escalation_f1','avg_composite_reward']:\n",
|
| 747 |
+
" b = baseline_results.get(k, 0)\n",
|
| 748 |
+
" t = trained_results.get(k, 0)\n",
|
| 749 |
+
" print(f' {k}: {b:.2f} β {t:.2f} ({t-b:+.2f})')\n",
|
| 750 |
+
"print()\n",
|
| 751 |
+
"print('π¦ Artifacts:')\n",
|
| 752 |
+
"print(f' LoRA weights: https://huggingface.co/{HF_REPO}')\n",
|
| 753 |
+
"print(f' HF Space: https://huggingface.co/spaces/{HF_USERNAME}/{SPACE_NAME}')\n",
|
| 754 |
+
"print(f' W&B run: {wandb.run.url if wandb.run else \"(finished)\"}')\n",
|
| 755 |
+
"print()\n",
|
| 756 |
+
"print('π Next steps:')\n",
|
| 757 |
+
"print(' 1. Fill in blog_draft.md with real numbers')\n",
|
| 758 |
+
"print(' 2. Publish blog at huggingface.co/new-blog')\n",
|
| 759 |
+
"print(' 3. Record 2 min video showing before/after')\n",
|
| 760 |
+
"print(' 4. Add blog + video URLs to README.md')\n",
|
| 761 |
+
"print(' 5. Submit HF Space URL to judges')"
|
| 762 |
+
],
|
| 763 |
+
"execution_count": null,
|
| 764 |
+
"outputs": []
|
| 765 |
+
}
|
| 766 |
+
|
| 767 |
+
]
|
| 768 |
+
}
|