{ "torch": "2.15.0+cu132", "fixture_sha256": "aff0d4c84ff104219bf365a13dcb0de1c1cfc59b833ac72cb357c46e29a83f13", "fixture": { "grpo": { "loss_actual": 429.46087646484375, "loss_reference": 429.46087646484375, "loss_error": { "max_abs": 0.0, "exact": true }, "gradient_error": { "max_abs": 0.0, "exact": true }, "diagnostics": { "n_loss_tokens": 1024.0, "n_nonzero_adv_tokens": 992, "dppo_removed_gradient_tokens": 183, "dppo_masked_nonzero_adv_fraction": 0.1844758064516129, "policy_drift_abs_p_q99": 0.6322206427156923, "policy_drift_abs_p_max": 0.7907273918390274, "policy_drift_signed_p_min": -0.7907273918390274, "policy_drift_signed_p_max": 0.7162230312824249, "abs_drift_above_delta_tokens": 353, "dppo_masked_token_fraction": 0.1787109375, "grpo_clip_fraction": 0.3525390625, "policy_drift_abs_p_q90": 0.35204049944877625, "policy_drift_abs_p_mean": 0.13980250578968256, "mean_ratio": 1.0783979892730713, "nonzero_adv_fraction": 0.96875 } }, "dppo": { "loss_actual": -420.5429992675781, "loss_reference": -420.54302978515625, "loss_error": { "max_abs": 3.0517578125e-05, "exact": false }, "gradient_error": { "max_abs": 0.0, "exact": true }, "diagnostics": { "n_loss_tokens": 1024.0, "n_nonzero_adv_tokens": 992, "dppo_removed_gradient_tokens": 183, "dppo_masked_nonzero_adv_fraction": 0.1844758064516129, "policy_drift_abs_p_q99": 0.6322206427156923, "policy_drift_abs_p_max": 0.7907273918390274, "policy_drift_signed_p_min": -0.7907273918390274, "policy_drift_signed_p_max": 0.7162230312824249, "abs_drift_above_delta_tokens": 353, "dppo_masked_token_fraction": 0.1787109375, "grpo_clip_fraction": 0.3525390625, "policy_drift_abs_p_q90": 0.35204049944877625, "policy_drift_abs_p_mean": 0.13980250578968256, "mean_ratio": 1.0783979892730713, "nonzero_adv_fraction": 0.96875 } }, "mask_exact": true }, "deterministic": [ { "case": "positive_block", "gradient": -0.0, "expected_gradient": 0.0, "reference_keep": 0.0 }, { "case": "negative_block", "gradient": 0.0, "expected_gradient": 0.0, "reference_keep": 0.0 }, { "case": "positive_recovery", "gradient": -0.6666666666666666, "expected_gradient": -0.6666666666666666, "reference_keep": 1.0 }, { "case": "negative_recovery", "gradient": 1.5, "expected_gradient": 1.5, "reference_keep": 1.0 }, { "case": "positive_small", "gradient": -1.2, "expected_gradient": -1.2, "reference_keep": 1.0 }, { "case": "negative_small", "gradient": 0.8333333333333334, "expected_gradient": 0.8333333333333334, "reference_keep": 1.0 }, { "case": "tail_positive", "gradient": -10.0, "expected_gradient": -10.0, "reference_keep": 1.0 }, { "case": "tail_negative", "gradient": 9.999999999999973e-05, "expected_gradient": 9.999999999999999e-05, "reference_keep": 1.0 }, { "case": "cap_positive", "gradient": -10.0, "expected_gradient": -10.0, "reference_keep": 1.0 }, { "case": "zero_adv", "gradient": 0.0, "expected_gradient": 0.0, "reference_keep": 0.0 }, { "case": "padding", "gradient": -0.0, "expected_gradient": 0.0, "reference_keep": 0.0 } ], "deterministic_ppo_gradient_and_clip_diagnostic": true, "default_delta_both_sides_both_signs": true, "strict_boundary_both_signs": true, "zero_adv_diagnostic": { "n_loss_tokens": 1.0, "n_nonzero_adv_tokens": 0, "dppo_removed_gradient_tokens": 0, "dppo_masked_nonzero_adv_fraction": null, "policy_drift_abs_p_q99": 0.25, "policy_drift_abs_p_max": 0.25, "policy_drift_signed_p_min": -0.25, "policy_drift_signed_p_max": -0.25, "abs_drift_above_delta_tokens": 1, "dppo_masked_token_fraction": 0.0, "grpo_clip_fraction": 0.0, "policy_drift_abs_p_q90": 0.25, "policy_drift_abs_p_mean": 0.25, "mean_ratio": 0.6666666269302368, "nonzero_adv_fraction": 0.0 }, "pilot": { "tokens": 430133.0, "per_update": [ { "update": 1, "tokens": 211980.0, "masked_tokens": 0.0, "ppo_clipped_tokens": 79, "q90_range": [ 0.008316918276250362, 0.011326874606311321 ], "token_weighted_abs_drift_mean": 0.0027594480924321833 }, { "update": 2, "tokens": 218153.0, "masked_tokens": 0.0, "ppo_clipped_tokens": 100, "q90_range": [ 0.009953003376722336, 0.011339953169226646 ], "token_weighted_abs_drift_mean": 0.0029434709941956734 } ], "raw_token_tails_available": false } }