File size: 6,646 Bytes
c17c7eb | 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 | """
AegisGym GRPO Training Script
Connects to the live HF Space for environment rollouts.
Run: python train.py
"""
import os
import json
from datasets import Dataset
from trl import GRPOConfig, GRPOTrainer
from transformers import AutoTokenizer
from client_env import get_sync_client
from models import AuditAction
# ─── Config ──────────────────────────────────────────────────────────────────
ENV_URL = os.getenv("AEGISGYM_URL", "https://armaan020-aegisgym.hf.space")
MODEL_NAME = os.getenv("MODEL_NAME", "Qwen/Qwen2.5-0.5B-Instruct")
OUTPUT_DIR = "aegisgym-grpo-agent"
SYSTEM_PROMPT = """You are a financial compliance auditor AI.
Given a transaction scenario and regulatory context, respond with a JSON object:
{"action_type": "FLAG|APPROVE|REQUEST_INFO", "target_id": "<account_id>", "regulation_citation": "<regulation>"}
Be precise and concise."""
# ─── Action parser ────────────────────────────────────────────────────────────
def parse_action(text: str) -> AuditAction:
try:
start = text.find("{")
end = text.rfind("}") + 1
if start >= 0 and end > start:
data = json.loads(text[start:end])
return AuditAction(**data)
except Exception:
pass
return AuditAction(action_type="REQUEST_INFO", target_id="UNKNOWN", regulation_citation="parse_error")
# ─── Dataset Loading ──────────────────────────────────────────────────────────
from datasets import load_dataset
import itertools
print("Loading SecureFinAI-Lab/Regulations_QA dataset...")
try:
# Use a streaming dataset for efficiency
raw_dataset = load_dataset("SecureFinAI-Lab/Regulations_QA", split="train", streaming=True)
dataset_iterator = itertools.cycle(iter(raw_dataset))
print("Dataset loaded successfully.")
except Exception as e:
print(f"Warning: Failed to load dataset: {e}. Falling back to default prompts.")
dataset_iterator = None
# ─── Rollout function ─────────────────────────────────────────────────────────
def rollout_func(trainer, prompts, tokenizer):
"""One rollout episode connecting to the live AegisGym Space with dataset augmentation."""
from trl.experimental.openenv import generate_rollout_completions
env = get_sync_client(ENV_URL)
result = env.reset()
obs_dict = result.get("observation", {})
# Sample from dataset if available
dataset_context = ""
custom_prompt = "Audit the following transaction."
if dataset_iterator:
item = next(dataset_iterator)
custom_prompt = item.get("question", custom_prompt)
dataset_context = f"\nRegulatory Context: {item.get('answer', '')}"
all_prompt_ids, all_completion_ids, all_logprobs, rewards = [], [], [], []
state = env.state()
tier = state.get("current_tier", "easy")
# Combine dataset context with environment context
regs = obs_dict.get('retrieved_regs', "")
retrieved_regs = [regs] if isinstance(regs, str) else list(regs)
if dataset_context:
retrieved_regs.append(dataset_context)
user_msg = (
f"{custom_prompt}\n\n"
f"Tier: {tier.upper()}\n"
f"Transactions: {obs_dict.get('transactions', [])}\n"
f"Context: {retrieved_regs}\n"
f"Account: {obs_dict.get('account_metadata', {})}"
)
messages = [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": user_msg},
]
prompt_text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
out = generate_rollout_completions(trainer, [prompt_text])[0]
all_prompt_ids.extend(out["prompt_ids"])
all_completion_ids.extend(out["completion_ids"])
all_logprobs.extend(out["logprobs"])
completion_text = out.get("text") or tokenizer.decode(out["completion_ids"], skip_special_tokens=True)
action = parse_action(completion_text)
result = env.step(action.model_dump())
rewards.append(float(result.get("reward", 0.0)))
return {
"prompt_ids": all_prompt_ids,
"completion_ids": all_completion_ids,
"logprobs": all_logprobs,
"env_reward": rewards[-1] if rewards else 0.0,
}
# ─── Reward shim ─────────────────────────────────────────────────────────────
def reward_compliance(completions, **kwargs):
rewards = kwargs.get("env_reward", [])
if not rewards:
return [0.0] * len(completions)
return [float(r) for r in rewards]
# ─── Main ─────────────────────────────────────────────────────────────────────
def main():
print(f"Model : {MODEL_NAME}")
print(f"Env : {ENV_URL}")
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
tokenizer.pad_token = tokenizer.eos_token
dataset = Dataset.from_dict({"prompt": ["Audit the following transaction."] * 200})
config = GRPOConfig(
num_train_epochs = 1,
learning_rate = 5e-6,
per_device_train_batch_size = 1,
gradient_accumulation_steps = 4,
warmup_steps = 10,
num_generations = 2,
max_completion_length = 256,
use_vllm = False,
output_dir = OUTPUT_DIR,
logging_steps = 1,
save_steps = 25,
gradient_checkpointing = True,
)
trainer = GRPOTrainer(
model = MODEL_NAME,
processing_class = tokenizer,
reward_funcs = [reward_compliance],
train_dataset = dataset,
args = config,
rollout_func = rollout_func,
)
print("\n=== Starting GRPO Training ===")
trainer.train()
trainer.save_model(OUTPUT_DIR)
print(f"\nModel saved to {OUTPUT_DIR}/")
if __name__ == "__main__":
main()
|