| |
| """Train GRPO policy with TRL using environment-backed verifier rewards.""" |
|
|
| from __future__ import annotations |
|
|
| import argparse |
| import json |
| from pathlib import Path |
|
|
| import sys |
|
|
| ROOT = Path(__file__).resolve().parents[1] |
| if str(ROOT) not in sys.path: |
| sys.path.insert(0, str(ROOT)) |
|
|
| from app.training.grpo_trl import GRPOTrlConfig, run_grpo_trl |
|
|
|
|
| def parse_args() -> argparse.Namespace: |
| parser = argparse.ArgumentParser(description="Run TRL GRPO with env-backed rewards.") |
| parser.add_argument("--model-id", default="Qwen/Qwen2.5-1.5B-Instruct") |
| parser.add_argument("--prompts-path", default="data/processed/training_corpus_grpo_prompts.jsonl") |
| parser.add_argument("--output-dir", default="checkpoints") |
| parser.add_argument("--report-path", default="outputs/reports/grpo_trl_run.json") |
| parser.add_argument("--max-prompts", type=int, default=256) |
| parser.add_argument("--max-steps", type=int, default=30) |
| parser.add_argument("--epochs", type=float, default=1.0) |
| parser.add_argument("--episodes", type=int, default=0, help="Backward-compatible alias for --max-steps.") |
| parser.add_argument("--batch-size", type=int, default=2) |
| parser.add_argument("--grad-accum", type=int, default=1) |
| parser.add_argument("--num-generations", type=int, default=2) |
| parser.add_argument("--max-prompt-length", type=int, default=512) |
| parser.add_argument("--max-completion-length", type=int, default=96) |
| parser.add_argument("--learning-rate", type=float, default=1e-6) |
| parser.add_argument("--temperature", type=float, default=0.7) |
| parser.add_argument("--seed", type=int, default=42) |
| parser.add_argument("--use-unsloth", action="store_true") |
| parser.add_argument("--allow-fallback", action="store_true") |
| parser.add_argument("--force-fallback", action="store_true") |
| return parser.parse_args() |
|
|
|
|
| def main() -> None: |
| args = parse_args() |
| root = Path(__file__).resolve().parents[1] |
|
|
| cfg = GRPOTrlConfig( |
| model_id=args.model_id, |
| prompts_path=root / args.prompts_path, |
| output_dir=root / args.output_dir, |
| max_prompts=args.max_prompts, |
| max_steps=args.episodes if args.episodes > 0 else args.max_steps, |
| epochs=args.epochs, |
| per_device_batch_size=args.batch_size, |
| gradient_accumulation_steps=args.grad_accum, |
| num_generations=args.num_generations, |
| learning_rate=args.learning_rate, |
| max_prompt_length=args.max_prompt_length, |
| max_completion_length=args.max_completion_length, |
| temperature=args.temperature, |
| seed=args.seed, |
| use_unsloth=args.use_unsloth, |
| allow_fallback=args.allow_fallback, |
| force_fallback=args.force_fallback, |
| ) |
|
|
| result = run_grpo_trl(cfg) |
|
|
| report_path = root / args.report_path |
| report_path.parent.mkdir(parents=True, exist_ok=True) |
| report_path.write_text(json.dumps(result, ensure_ascii=True, indent=2), encoding="utf-8") |
| print("grpo_trl_done") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
