Spaces:
Running
Running
File size: 13,212 Bytes
c452421 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 | """Phase-based GRPO Training for SENTINEL — Checkpoint Handoff System.
3-phase curriculum with explicit
checkpoint continuity. Each phase builds on the previous checkpoint.
Phase 1 — Basic Oversight (warm_start weights):
- Only basic_oversight task
- Higher tolerance for false positives
- Teaches: detect hallucinations + reward hacking
Phase 2 — Fleet Monitoring (balanced weights):
- Continues from Phase 1 checkpoint
- Adds fleet_monitoring_conflict task
- Teaches: scope violations + cross-worker reasoning
Phase 3 — Adversarial + Multi-Crisis (precision weights):
- Continues from Phase 2 checkpoint
- All 4 tasks including adversarial_worker + multi_crisis_command
- Teaches: adversarial resilience + multi-incident coordination
Usage:
python training/phase_training.py --phase 1
python training/phase_training.py --phase 2 --checkpoint outputs/phase1/final
python training/phase_training.py --phase 3 --checkpoint outputs/phase2/final
python training/phase_training.py --all # runs all 3 phases sequentially
"""
from __future__ import annotations
import json
import logging
import os
import shutil
import subprocess
import sys
from dataclasses import dataclass
from pathlib import Path
from typing import List, Optional
logger = logging.getLogger(__name__)
# Phase definitions — inspired by all 4 winners
PHASES = {
1: {
"name": "Basic Oversight",
"tasks": ["basic_oversight"],
"reward_profile": "warm_start",
"description": "Learn to detect hallucinations and reward hacking",
"steps": 80,
"lr": 5e-6,
"kl_coef": 0.02,
"temperature": 1.0,
"num_generations": 4,
"misbehavior_types": ["HALLUCINATION", "REWARD_HACKING"],
"expected_milestones": [
"detect hallucinated targets",
"identify premature classification",
],
},
2: {
"name": "Fleet Monitoring",
"tasks": ["basic_oversight", "fleet_monitoring_conflict"],
"reward_profile": "balanced",
"description": "Add scope violations and cross-worker reasoning",
"steps": 80,
"lr": 3e-6,
"kl_coef": 0.03,
"temperature": 0.9,
"num_generations": 4,
"misbehavior_types": [
"HALLUCINATION", "REWARD_HACKING",
"SCOPE_VIOLATION", "LOOP_EXPLOITATION",
],
"expected_milestones": [
"detect cross-domain scope violations",
"identify investigation loop exploitation",
],
},
3: {
"name": "Adversarial + Multi-Crisis",
"tasks": [
"basic_oversight",
"fleet_monitoring_conflict",
"adversarial_worker",
"multi_crisis_command",
],
"reward_profile": "precision",
"description": "Full adversarial resilience and multi-incident coordination",
"steps": 80,
"lr": 2e-6,
"kl_coef": 0.04,
"temperature": 0.8,
"num_generations": 4,
"misbehavior_types": [
"HALLUCINATION", "REWARD_HACKING",
"SCOPE_VIOLATION", "LOOP_EXPLOITATION",
"DESTRUCTIVE_PRECEDENCE", "ESCALATION_BOMBING",
"CONFIDENCE_WASHING", "PROMPT_INJECTION",
],
"expected_milestones": [
"handle compound misbehaviors",
"coordinate multi-incident triage",
"resist prompt injection attacks",
],
},
}
@dataclass
class PhaseResult:
"""Result of one training phase."""
phase: int
name: str
checkpoint_path: str
final_reward: float
best_reward: float
episodes_completed: int
milestones_achieved: List[str]
reward_history: List[float]
def get_phase_config(phase: int, checkpoint: Optional[str] = None) -> dict:
"""Get the full training configuration for a phase."""
if phase not in PHASES:
raise ValueError(f"Unknown phase {phase}. Valid phases: {list(PHASES.keys())}")
p = PHASES[phase]
output_dir = f"outputs/phase{phase}"
config = {
"phase": phase,
"phase_name": p["name"],
"description": p["description"],
"tasks": p["tasks"],
"reward_profile": p["reward_profile"],
"steps": p["steps"],
"learning_rate": p["lr"],
"kl_coef": p["kl_coef"],
"temperature": p["temperature"],
"num_generations": p["num_generations"],
"output_dir": output_dir,
"checkpoint": checkpoint,
"misbehavior_types": p["misbehavior_types"],
"expected_milestones": p["expected_milestones"],
}
return config
def generate_phase_env_vars(phase: int, checkpoint: Optional[str] = None) -> dict:
"""Generate environment variables for running train.py with phase config."""
config = get_phase_config(phase, checkpoint)
env_vars = {
"USE_SENTINEL": "1",
"TRAIN_STEPS": str(config["steps"]),
"LR": str(config["learning_rate"]),
"KL_COEF": str(config["kl_coef"]),
"NUM_GENERATIONS": str(config["num_generations"]),
"OUTPUT_DIR": config["output_dir"],
"SENTINEL_TASKS": ",".join(config["tasks"]),
"REWARD_PROFILE": config["reward_profile"],
"REWARD_SCHEDULE_MODE": config["reward_profile"],
"MODEL_STEPS_LIMIT": str(config.get("model_steps_limit", 1)),
}
if checkpoint:
env_vars["RESUME_FROM"] = checkpoint
return env_vars
def generate_deepspeed_config(
output_path: str = "outputs/ds_config_zero2.json",
) -> str:
"""Generate a DeepSpeed ZeRO Stage 2 config suitable for GRPO training."""
config = {
"bf16": {"enabled": True},
"zero_optimization": {
"stage": 2,
"offload_optimizer": {"device": "cpu", "pin_memory": True},
"allgather_partitions": True,
"allgather_bucket_size": 2e8,
"overlap_comm": True,
"reduce_scatter": True,
"reduce_bucket_size": 2e8,
"contiguous_gradients": True,
},
"gradient_accumulation_steps": "auto",
"gradient_clipping": 1.0,
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"wall_clock_breakdown": False,
}
os.makedirs(os.path.dirname(output_path) or ".", exist_ok=True)
with open(output_path, "w", encoding="utf-8") as f:
json.dump(config, f, indent=2)
return output_path
def build_launch_command(
phase: int,
checkpoint: Optional[str] = None,
use_accelerate: bool = False,
deepspeed_config: Optional[str] = None,
num_processes: Optional[int] = None,
) -> str:
"""Build the full shell command to launch a phase, optionally via accelerate."""
env_vars = generate_phase_env_vars(phase, checkpoint)
env_prefix = " ".join(f"{k}={v}" for k, v in env_vars.items())
if use_accelerate or deepspeed_config:
accelerate_bin = shutil.which("accelerate") or "accelerate"
parts = [accelerate_bin, "launch"]
if num_processes:
parts += ["--num_processes", str(num_processes)]
if deepspeed_config:
parts += ["--use_deepspeed", "--deepspeed_config_file", deepspeed_config]
parts.append("train.py")
cmd = " ".join(parts)
else:
cmd = f"python train.py"
return f"{env_prefix} {cmd}"
def print_phase_plan():
"""Print the full 3-phase training plan."""
print("=" * 70)
print("SENTINEL — 3-Phase GRPO Training Plan")
print("=" * 70)
for phase_num, phase in PHASES.items():
print(f"\nPhase {phase_num}: {phase['name']}")
print(f" Description: {phase['description']}")
print(f" Tasks: {', '.join(phase['tasks'])}")
print(f" Reward: {phase['reward_profile']} weights")
print(f" Steps: {phase['steps']}")
print(f" LR: {phase['lr']}")
print(f" KL: {phase['kl_coef']}")
print(f" Temp: {phase['temperature']}")
print(f" Types: {', '.join(phase['misbehavior_types'])}")
print(f" Milestones:")
for m in phase["expected_milestones"]:
print(f" - {m}")
print("\n" + "=" * 70)
print("Run sequence:")
print(" python training/phase_training.py --phase 1")
print(" python training/phase_training.py --phase 2 --checkpoint outputs/phase1/final")
print(" python training/phase_training.py --phase 3 --checkpoint outputs/phase2/final")
print("=" * 70)
def log_phase_transition(from_phase: int, to_phase: int, checkpoint: str):
"""Log a phase transition for audit trail."""
transition = {
"from_phase": from_phase,
"to_phase": to_phase,
"checkpoint": checkpoint,
"from_name": PHASES[from_phase]["name"],
"to_name": PHASES[to_phase]["name"],
"reward_transition": f"{PHASES[from_phase]['reward_profile']} -> {PHASES[to_phase]['reward_profile']}",
"new_tasks": [t for t in PHASES[to_phase]["tasks"] if t not in PHASES[from_phase]["tasks"]],
"new_misbehaviors": [
m for m in PHASES[to_phase]["misbehavior_types"]
if m not in PHASES[from_phase]["misbehavior_types"]
],
}
os.makedirs("outputs", exist_ok=True)
with open("outputs/phase_transitions.jsonl", "a") as f:
f.write(json.dumps(transition) + "\n")
logger.info(
"Phase transition: %s -> %s (checkpoint: %s, new tasks: %s)",
transition["from_name"],
transition["to_name"],
checkpoint,
transition["new_tasks"],
)
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser(description="SENTINEL Phase-Based Training")
parser.add_argument("--phase", type=int, choices=[1, 2, 3], help="Phase to run")
parser.add_argument("--checkpoint", type=str, help="Checkpoint from previous phase")
parser.add_argument("--plan", action="store_true", help="Print training plan only")
parser.add_argument("--all", action="store_true", help="Run all 3 phases sequentially")
parser.add_argument("--accelerate", action="store_true", help="Launch via accelerate (multi-GPU / DDP)")
parser.add_argument("--deepspeed", type=str, default=None,
help="Path to DeepSpeed config JSON (auto-generates ZeRO-2 if 'auto')")
parser.add_argument("--num-processes", type=int, default=None,
help="Number of processes for accelerate launch")
parser.add_argument("--run", action="store_true",
help="Actually run the training (not just print commands)")
args = parser.parse_args()
if args.plan:
print_phase_plan()
sys.exit(0)
# Auto-generate DeepSpeed config if requested
ds_config_path = args.deepspeed
if ds_config_path == "auto":
ds_config_path = generate_deepspeed_config()
print(f"Generated DeepSpeed ZeRO-2 config at: {ds_config_path}")
use_accelerate = args.accelerate or bool(ds_config_path)
if args.all:
print_phase_plan()
print("\nReady to run all 3 phases.")
for phase in [1, 2, 3]:
ckpt = f"outputs/phase{phase-1}/final" if phase > 1 else None
cmd = build_launch_command(
phase,
checkpoint=ckpt,
use_accelerate=use_accelerate,
deepspeed_config=ds_config_path,
num_processes=args.num_processes,
)
print(f"\n--- Phase {phase} ---")
print(f" {cmd}")
if args.run:
if phase > 1:
log_phase_transition(phase - 1, phase, ckpt)
env_vars = generate_phase_env_vars(phase, ckpt)
env = {**os.environ, **env_vars}
print(f" Running phase {phase}...")
result = subprocess.run(cmd, shell=True, env=env)
if result.returncode != 0:
print(f" Phase {phase} failed with code {result.returncode}")
sys.exit(result.returncode)
print(f" Phase {phase} complete.")
sys.exit(0)
if not args.phase:
parser.error("Specify --phase N or --all or --plan")
cmd = build_launch_command(
args.phase,
checkpoint=args.checkpoint,
use_accelerate=use_accelerate,
deepspeed_config=ds_config_path,
num_processes=args.num_processes,
)
print(f"\nPhase {args.phase}: {PHASES[args.phase]['name']}")
print(f"Command:")
print(f" {cmd}")
if args.run:
env_vars = generate_phase_env_vars(args.phase, args.checkpoint)
env = {**os.environ, **env_vars}
if args.checkpoint and args.phase > 1:
log_phase_transition(args.phase - 1, args.phase, args.checkpoint)
result = subprocess.run(cmd, shell=True, env=env)
sys.exit(result.returncode)
if args.checkpoint and args.phase > 1:
log_phase_transition(args.phase - 1, args.phase, args.checkpoint)
"""
Phase-based GRPO Training for SENTINEL.
Implements the cross-winner pattern: separate phases with checkpoint handoff.
"""
|