| """Resolve the pinned base checkpoint and create a reproducible training config.""" |
|
|
| import argparse |
| import json |
| from pathlib import Path |
| from huggingface_hub import snapshot_download |
|
|
| parser = argparse.ArgumentParser(description=__doc__) |
| parser.add_argument("--seed", type=int, choices=range(5), required=True) |
| parser.add_argument("--device", default="cuda:0") |
| parser.add_argument("--out", required=True) |
| args = parser.parse_args() |
| root = Path(__file__).resolve().parent |
| metadata = json.loads((root / "adapter_config.json").read_text()) |
| base = snapshot_download( |
| metadata["base_model"], |
| revision=metadata["base_revision"], |
| allow_patterns=["model.safetensors", "rl_agent_config.json", "encoder/*", "tokenizer/*"], |
| ) |
| config = { |
| "name": f"Deterministic frozen base + linear, seed {args.seed} (UNTRAINED)", |
| "adapter": "python", |
| "mode": "reference", |
| "factory": "ariadne_bench.frozen_input_interface:create", |
| "options": { |
| "laya_model": str(Path(base).resolve()), |
| "device": args.device, |
| "batch_size": 16, |
| "max_len": 1024, |
| "head_max_len": 256, |
| "seed": args.seed, |
| "training_scope": "bridge", |
| "add_interface": True, |
| "deterministic": True, |
| }, |
| } |
| Path(args.out).write_text(json.dumps(config, indent=2) + "\n") |
|
|
