File size: 10,510 Bytes
aec0295 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 | """
GRPO training entrypoint for the local trading policy.
This script is intended for GPU-backed Hugging Face or local Linux runs where
Unsloth is available. It uses the same prompt schema as the runtime policy and
the verifier functions in env.reward.
"""
from __future__ import annotations
import os
os.environ.setdefault("OPENBLAS_NUM_THREADS", "1")
os.environ.setdefault("OMP_NUM_THREADS", "1")
import argparse
import inspect
import json
import random
import sys
from pathlib import Path
import numpy as np
from datasets import Dataset
ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
from env.reward import (
alignment_reward_func,
format_reward_func,
governance_reward_func,
profit_reward_func,
risk_reward_func,
)
from utils.plotting import plot_training_results
DEFAULT_MODEL_NAME = "unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit"
DEFAULT_OUTPUT_DIR = "models/local_policy_grpo"
DEFAULT_TRAJECTORY_PATH = "checkpoints/sft_trajectories.jsonl"
SYSTEM_PROMPT = """You are a Quant Trader operating inside a multi-agent market simulation.
Read the JSON scenario carefully and produce exactly one action.
Respond exactly in this format:
<thought>
Short reasoning about trend, fundamentals, and risk.
</thought>
<action>
{"direction": 0, "size": 0.0}
</action>
"""
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="Train the trading policy with GRPO.")
parser.add_argument("--model-name", default=DEFAULT_MODEL_NAME)
parser.add_argument("--output-dir", default=DEFAULT_OUTPUT_DIR)
parser.add_argument("--trajectory-path", default=DEFAULT_TRAJECTORY_PATH)
parser.add_argument("--regime", choices=["easy", "medium", "hard"], default="easy")
parser.add_argument("--max-seq-length", type=int, default=1024)
parser.add_argument("--max-prompt-length", type=int, default=768)
parser.add_argument("--max-completion-length", type=int, default=200)
parser.add_argument("--max-steps", type=int, default=250)
parser.add_argument("--save-steps", type=int, default=50)
parser.add_argument("--logging-steps", type=int, default=1)
parser.add_argument("--per-device-batch-size", type=int, default=4)
parser.add_argument("--gradient-accumulation-steps", type=int, default=2)
parser.add_argument("--num-generations", type=int, default=4)
parser.add_argument("--learning-rate", type=float, default=5e-5)
parser.add_argument("--min-grade", type=float, default=0.65)
parser.add_argument("--max-records", type=int, default=512)
parser.add_argument("--num-scenarios", type=int, default=500)
parser.add_argument("--seed", type=int, default=3407)
return parser.parse_args()
def build_prompt(state: list[float], signals: dict[str, float]) -> str:
scenario = {
"state": state,
"signals": {
"ta": float(signals["ta"]),
"fa": float(signals["fa"]),
"position_limit": float(signals["position_limit"]),
},
}
return f"{SYSTEM_PROMPT}\nScenario:\n{json.dumps(scenario, separators=(',', ':'))}\n"
def synthetic_scenarios(regime: str, n: int = 500) -> list[dict]:
"""Generate *n* diverse synthetic market scenarios.
Each scenario has a short price-state snippet (5 ticks) and
randomized TA/FA signals with a position limit. The regime
biases the distribution so curriculum learning works:
easy — mostly trending, clear signals
medium — mixed, some conflicting signals
hard — high vol, noisy & contradictory signals
"""
rng = np.random.default_rng()
samples: list[dict] = []
for _ in range(n):
# --- price snippet (5 ticks, normalized around 1.0) ---
if regime == "easy":
trend = rng.choice([0.01, -0.01]) # clear up or down
noise = rng.normal(0, 0.005, 5)
elif regime == "medium":
trend = rng.normal(0, 0.005) # weak trend
noise = rng.normal(0, 0.01, 5)
else:
trend = rng.normal(0, 0.01) # ambiguous
noise = rng.normal(0, 0.03, 5)
base = 1.0
state = [round(base + trend * i + noise[i], 4) for i in range(5)]
# --- signals ---
is_up = state[-1] > state[0]
if regime == "easy":
# TA strongly agrees with trend
ta = rng.uniform(0.5, 1.0) if is_up else rng.uniform(-1.0, -0.5)
fa = rng.uniform(-0.3, 0.5) if is_up else rng.uniform(-0.5, 0.3)
elif regime == "medium":
ta = rng.uniform(-0.5, 0.5) # ambiguous
fa = rng.uniform(-0.5, 0.5)
else:
# Signals may contradict the trend
ta = rng.uniform(-1.0, 1.0)
fa = rng.uniform(-1.0, 1.0)
position_limit = float(rng.choice([0.2, 0.3, 0.5, 0.7, 0.8, 1.0]))
samples.append({
"state": state,
"signals": {
"ta": round(float(ta), 3),
"fa": round(float(fa), 3),
"position_limit": position_limit,
},
})
return samples
def load_trajectory_scenarios(path: str, min_grade: float, max_records: int) -> list[dict]:
if not os.path.exists(path):
return []
records: list[dict] = []
with open(path, "r", encoding="utf-8") as handle:
for line in handle:
row = json.loads(line)
if row.get("final_grade", 0.0) < min_grade:
continue
signal_blob = row.get("signals", {})
records.append(
{
"state": [float(x) for x in row.get("state", [])],
"signals": {
"ta": float(signal_blob.get("ta_score", 0.0)),
"fa": float(signal_blob.get("fa_sentiment", 0.0)),
"position_limit": float(signal_blob.get("position_limit", 1.0)),
},
}
)
random.shuffle(records)
return records[:max_records]
def build_dataset(args: argparse.Namespace) -> Dataset:
random.seed(args.seed)
scenarios = load_trajectory_scenarios(
path=args.trajectory_path,
min_grade=args.min_grade,
max_records=args.max_records,
)
if not scenarios:
scenarios = synthetic_scenarios(args.regime, n=args.num_scenarios)
prompts = [{"prompt": build_prompt(item["state"], item["signals"])} for item in scenarios]
return Dataset.from_list(prompts)
def require_cuda():
import torch
if not torch.cuda.is_available():
raise SystemExit(
"GRPO training requires CUDA. Unsloth does not support CPU-only execution."
)
return torch
def load_model(model_name: str, max_seq_length: int):
from unsloth import FastLanguageModel, PatchFastRL
PatchFastRL("GRPO", "unsloth")
model, tokenizer = FastLanguageModel.from_pretrained(
model_name=model_name,
max_seq_length=max_seq_length,
dtype=None,
load_in_4bit=True,
)
model = FastLanguageModel.get_peft_model(
model,
r=16,
target_modules=[
"q_proj",
"k_proj",
"v_proj",
"o_proj",
"gate_proj",
"up_proj",
"down_proj",
],
lora_alpha=16,
lora_dropout=0,
bias="none",
use_gradient_checkpointing="unsloth", # type: ignore
random_state=3407,
use_rslora=False,
loftq_config=None,
)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
return model, tokenizer
def make_trainer(model, tokenizer, dataset: Dataset, args: argparse.Namespace, torch_module):
from trl.trainer.grpo_config import GRPOConfig
from trl.trainer.grpo_trainer import GRPOTrainer
training_args = GRPOConfig(
output_dir=args.output_dir,
learning_rate=args.learning_rate,
per_device_train_batch_size=args.per_device_batch_size,
gradient_accumulation_steps=args.gradient_accumulation_steps,
num_train_epochs=1,
max_steps=args.max_steps,
save_steps=args.save_steps,
logging_steps=args.logging_steps,
bf16=torch_module.cuda.is_bf16_supported(),
fp16=not torch_module.cuda.is_bf16_supported(),
max_prompt_length=args.max_prompt_length, # type: ignore
max_completion_length=args.max_completion_length,
num_generations=args.num_generations,
report_to="none",
)
trainer_kwargs = {
"model": model,
"reward_funcs": [
format_reward_func,
alignment_reward_func,
risk_reward_func,
profit_reward_func,
governance_reward_func,
],
"args": training_args,
"train_dataset": dataset,
}
trainer_signature = inspect.signature(GRPOTrainer.__init__)
if "processing_class" in trainer_signature.parameters:
trainer_kwargs["processing_class"] = tokenizer
elif "tokenizer" in trainer_signature.parameters:
trainer_kwargs["tokenizer"] = tokenizer
return GRPOTrainer(**trainer_kwargs)
def save_model(model, tokenizer, output_dir: str) -> None:
os.makedirs(output_dir, exist_ok=True)
if hasattr(model, "save_pretrained_merged"):
model.save_pretrained_merged(output_dir, tokenizer, save_method="merged_16bit")
else:
model.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
def main() -> None:
args = parse_args()
torch_module = require_cuda()
dataset = build_dataset(args)
model, tokenizer = load_model(args.model_name, args.max_seq_length)
trainer = make_trainer(model, tokenizer, dataset, args, torch_module)
print(f"Starting GRPO training on {len(dataset)} prompts...")
train_result = trainer.train()
# Generate Plots
metrics = train_result.metrics
# TRL GRPOTrainer logs 'loss' and 'reward' in logs. We extract them from the history.
history = trainer.state.log_history
rewards = [x['reward'] for x in history if 'reward' in x]
losses = [x['loss'] for x in history if 'loss' in x]
plot_training_results(rewards, losses)
print(f"Saving GRPO policy to {args.output_dir}...")
save_model(model, tokenizer, args.output_dir)
print("GRPO training complete.")
if __name__ == "__main__":
main()
|