File size: 8,283 Bytes
ddbc1ba | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 | """
Compare base vs trained LifeStack policy on identical crisis prompts.
Usage:
python scripts/compare_baseline.py
python scripts/compare_baseline.py --trained-model ./lifestack_model
"""
import argparse
import json
import os
import random
import sys
from datetime import datetime
from typing import Any
import torch
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
REPO_ROOT = os.path.dirname(SCRIPT_DIR)
sys.path.insert(0, REPO_ROOT)
sys.path.insert(0, SCRIPT_DIR)
from agent.conflict_generator import TaskGenerator, generate_conflict
from core.life_state import DependencyGraph, LifeMetrics, ResourceBudget
from intake.simperson import SimPerson
from scripts.train_trl import build_prompt_for_task, get_lifestack_evaluation
def _load_base_model():
"""Load base Qwen2.5-1.5B-Instruct (no training adapter)."""
try:
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
model_name="unsloth/Qwen2.5-1.5B-Instruct",
max_seq_length=1024,
load_in_4bit=True,
)
FastLanguageModel.for_inference(model)
return model, tokenizer, "unsloth/base-qwen2.5-1.5b-instruct"
except Exception:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "Qwen/Qwen2.5-1.5B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
device_map="auto",
)
model.eval()
return model, tokenizer, model_name
def _load_trained_model(model_dir: str):
"""Load trained LifeStack model from local adapter/full checkpoint directory."""
try:
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
model_name=model_dir,
max_seq_length=1024,
load_in_4bit=True,
)
FastLanguageModel.for_inference(model)
return model, tokenizer
except Exception:
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(model_dir)
base = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen2.5-1.5B-Instruct",
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
device_map="auto",
)
model = PeftModel.from_pretrained(base, model_dir)
model.eval()
return model, tokenizer
def _device_for(model) -> torch.device:
try:
return next(model.parameters()).device
except Exception:
return torch.device("cuda" if torch.cuda.is_available() else "cpu")
def _generate_completion(model, tokenizer, prompt: str, temperature: float = 0.3) -> str:
device = _device_for(model)
inputs = tokenizer(prompt, return_tensors="pt").to(device)
pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=128,
temperature=temperature,
do_sample=True,
top_p=0.9,
pad_token_id=pad_token_id,
eos_token_id=tokenizer.eos_token_id,
)
return tokenizer.decode(outputs[0][inputs["input_ids"].shape[1] :], skip_special_tokens=True).strip()
def _build_eval_cases() -> list[dict[str, Any]]:
"""Create 5 deterministic prompts spanning different crisis domains."""
domains = [
("career", 3, 101),
("finances", 4, 202),
("relationships", 3, 303),
("transport_crisis", 4, 404),
("code_merge_crisis", 5, 505),
]
generator = TaskGenerator()
graph = DependencyGraph()
person = SimPerson(name="Comparator")
cases: list[dict[str, Any]] = []
for domain, difficulty, seed in domains:
random.seed(seed)
task = generator.generate(domain=domain, difficulty=difficulty)
conflict = generate_conflict(difficulty)
random.seed()
metrics = LifeMetrics()
metrics = graph.cascade(metrics, {**task.mutable_world, **conflict.primary_disruption})
budget_dict = task.constraints.get("budget", {})
budget = ResourceBudget(
time_hours=budget_dict.get("time", 20.0),
money_dollars=budget_dict.get("money", 500.0),
energy_units=budget_dict.get("energy", 100.0),
)
prompt = build_prompt_for_task(task, person, metrics, budget, seed=seed, step=0)
crisis_text = task.domain_metadata.get("story", task.goal)
cases.append(
{
"case_id": f"{domain}_d{difficulty}",
"domain": domain,
"difficulty": difficulty,
"seed": seed,
"crisis": crisis_text,
"prompt": prompt,
}
)
return cases
def _print_case(case: dict[str, Any]) -> None:
print("=" * 110)
print(f"[{case['case_id']}] domain={case['domain']} difficulty={case['difficulty']}")
print(f"crisis: {case['crisis']}")
print(f"base_reward={case['base_reward']:.3f} | trained_reward={case['trained_reward']:.3f} | delta={case['delta']:+.3f}")
print("- BASE RESPONSE -")
print(case["base_response"] or "<empty>")
print("- TRAINED RESPONSE -")
print(case["trained_response"] or "<empty>")
def run_compare(trained_model_dir: str, output_path: str) -> dict[str, Any]:
cases = _build_eval_cases()
print("Loading base model...")
base_model, base_tokenizer, base_name = _load_base_model()
for case in cases:
completion = _generate_completion(base_model, base_tokenizer, case["prompt"])
eval_data = get_lifestack_evaluation(completion, case["prompt"])
case["base_model"] = base_name
case["base_response"] = completion
case["base_reward"] = float(eval_data.get("reward", -0.5))
del base_model
torch.cuda.empty_cache()
print("Loading trained model...")
trained_model, trained_tokenizer = _load_trained_model(trained_model_dir)
for case in cases:
completion = _generate_completion(trained_model, trained_tokenizer, case["prompt"])
eval_data = get_lifestack_evaluation(completion, case["prompt"])
case["trained_model"] = trained_model_dir
case["trained_response"] = completion
case["trained_reward"] = float(eval_data.get("reward", -0.5))
case["delta"] = round(case["trained_reward"] - case["base_reward"], 4)
_print_case(case)
del trained_model
torch.cuda.empty_cache()
avg_base = sum(c["base_reward"] for c in cases) / len(cases)
avg_trained = sum(c["trained_reward"] for c in cases) / len(cases)
avg_delta = avg_trained - avg_base
payload = {
"timestamp_utc": datetime.utcnow().isoformat() + "Z",
"summary": {
"n_cases": len(cases),
"avg_base_reward": round(avg_base, 4),
"avg_trained_reward": round(avg_trained, 4),
"avg_reward_delta": round(avg_delta, 4),
"base_model": cases[0]["base_model"] if cases else "",
"trained_model": trained_model_dir,
},
"cases": cases,
}
output_dir = os.path.dirname(output_path)
if output_dir:
os.makedirs(output_dir, exist_ok=True)
with open(output_path, "w", encoding="utf-8") as f:
json.dump(payload, f, indent=2)
print("=" * 110)
print(
f"SUMMARY: avg_base={avg_base:.3f} | avg_trained={avg_trained:.3f} | "
f"avg_delta={avg_delta:+.3f}"
)
print(f"Saved comparison JSON: {output_path}")
return payload
def main():
parser = argparse.ArgumentParser(description="Compare baseline Qwen vs trained LifeStack model.")
parser.add_argument("--trained-model", type=str, default="./lifestack_model")
parser.add_argument("--output", type=str, default="./data/before_after_comparison.json")
args = parser.parse_args()
run_compare(trained_model_dir=args.trained_model, output_path=args.output)
if __name__ == "__main__":
main()
|