| |
| """Write L15 diag configs with SEPARATE promote vs backtrack gates. |
| |
| Naming: diag_L15_prom{F|CE}{thr}_bt{F|CE|NONE}{thr} |
| promF095 = promote on hop-accuracy (frontier) >= 0.95 |
| btCE090 = backtrack on both-arms score (ce_score) < 0.90 |
| btNONE = no backtracking |
| """ |
| from pathlib import Path |
|
|
| ROOT = Path(__file__).resolve().parents[1] / "args" |
|
|
| TEMPLATE = """\ |
| # L15 DIAG — SEPARATE gates |
| # PROMOTE (stage i -> i+1): {promote_metric} >= {promote_threshold} |
| # BACKTRACK (retrain earlier): {bt_desc} |
| project: coconut |
| save_path: ckpts |
| name: "{name}" |
| |
| only_eval: False |
| coconut: True |
| cot: False |
| no_thoughts: False |
| no_cot: False |
| |
| c_thought: 1 |
| max_latent_stage: 15 |
| pad_latent_to_max: True |
| |
| accuracy_staging: True |
| init_stage: 0 |
| promote_metric: {promote_metric} |
| promote_threshold: {promote_threshold} |
| promote_on_current_only: {promote_on_current_only} |
| epochs_per_stage: 25 |
| |
| backtrack: {backtrack} |
| backtrack_metric: {backtrack_metric} |
| backtrack_detect_threshold: {backtrack_detect_threshold} |
| remember_rate: 0.3 |
| revert_staging: False |
| |
| eval_every: 10 |
| log_every: 5 |
| perhop_val_samples: 256 |
| perhop_train_samples: 64 |
| eval_print_full: False |
| |
| backprop_depth: null |
| train_size: 50000 |
| save_only_improve: False |
| save_every: 200 |
| uniform_prob: 0.1 |
| model_id: configs/symbol-2layer-8head-768dim-L20.json |
| load_model_path: ckpts/star-coconut-L15-bfs-stage0-warm/checkpoint_99 |
| seed: 0 |
| resume: 0 |
| bf16: True |
| train_path: data/star_2arm_L15_train_fo_bfs.json |
| val_path: data/star_2arm_L15_valid_fo_bfs.json |
| reset_optimizer: False |
| batch_size_training: 128 |
| debug: False |
| gradient_accumulation_steps: 1 |
| num_epochs: 3000 |
| lr: !!float "1e-4" |
| grad_clip: !!float "1.0" |
| warmup_steps: 200 |
| weight_decay: 0.01 |
| bfs_variant: True |
| """ |
|
|
| |
| ARMS = [ |
| ("diag_L15_promF095_btCE050", "frontier", 0.95, True, "ce_score", 0.50, False), |
| ("diag_L15_promF095_btCE090", "frontier", 0.95, True, "ce_score", 0.90, False), |
| ("diag_L15_promF095_btCE095", "frontier", 0.95, True, "ce_score", 0.95, False), |
| ("diag_L15_promF099_btCE090", "frontier", 0.99, True, "ce_score", 0.90, False), |
| ("diag_L15_promF095_btF095", "frontier", 0.95, True, "frontier", 0.95, False), |
| ("diag_L15_promCE090_btCE090", "ce_score", 0.90, True, "ce_score", 0.90, False), |
| ("diag_L15_promF095_btNONE", "frontier", 0.95, False, "frontier", 0.95, True), |
| ("diag_L15_promF099_btCE050", "frontier", 0.99, True, "ce_score", 0.50, False), |
| ] |
|
|
|
|
| def main(): |
| for name, pm, pt, bt, bm, bt_thr, cur_only in ARMS: |
| bt_desc = "OFF" if not bt else f"{bm} < {bt_thr} triggers retrain" |
| text = TEMPLATE.format( |
| name=name, |
| promote_metric=pm, |
| promote_threshold=pt, |
| promote_on_current_only=str(cur_only), |
| backtrack=str(bt), |
| backtrack_metric=bm, |
| backtrack_detect_threshold=bt_thr, |
| bt_desc=bt_desc, |
| ) |
| path = ROOT / f"{name}.yaml" |
| path.write_text(text) |
| print("wrote", path.name) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
