File size: 5,120 Bytes
ad89351 | 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 | {
"description": "Controlled scaling experiment: 8M vs 46.5M parameter Diffusion-LM riddle solvers",
"dataset": {
"name": "Riddles \u2014 A Synthetic Riddle Dataset for NLP",
"source": "Kaggle",
"url": "https://www.kaggle.com/datasets/prajwaldongre/riddles-a-synthetic-riddle-dataset-for-nlp",
"license": "CC0",
"train_size": 232,
"heldout_size": 66,
"unique_total": 298,
"duplicates_removed": 702,
"split_method": "SHA-256 of normalized riddle text",
"answer_max_len": 4,
"vocab_source": "training set only"
},
"checkpoints": {
"phase3": {
"label": "Phase 3 Reconstructed",
"architecture": {
"params": null,
"T": 200,
"d_model": 256,
"n_layers": 4
},
"training": {
"steps": 500,
"lr": 0.0003,
"schedule": "constant",
"batch": 64,
"seed": 42,
"T": 200
},
"eval_heldout": {
"exact_match_K1_pct": 46.96969696969697,
"exact_match_K10_pct": 46.96969696969697,
"token_f1_K1": 0.7124098124098122,
"n": 66
},
"eval_train": {
"exact_match_K1_pct": 87.5,
"exact_match_K10_pct": 86.45833333333333,
"token_f1_K1": 0.9597222222222221,
"n": 192
},
"decoding_modes": {
"A_full_chain_em_pct": 0.5,
"B_pure_noise_1shot_em_pct": 0.494949494949495,
"C_gold_corrupted_1shot_em_pct": 0.5303030303030303
},
"signal_degradation": {
"raw_xt_tok_acc_at_t2_pct": 100.0,
"model_tok_acc_at_t2_pct": 76.6,
"degradation_ppt": 23.4
},
"state_drift": {
"t2_oracle_tok_acc_pct": 75.9,
"t2_free_tok_acc_pct": 72.0,
"drift_ppt": 3.91,
"oracle_flat_across_noise": false
}
},
"phase4_original": {
"label": "Phase 4 Original",
"architecture": {
"params": null,
"T": 500,
"d_model": 512,
"n_layers": 6
},
"training": {
"steps": 20000,
"lr": 0.0003,
"schedule": "constant",
"batch": 64,
"seed": 42,
"T": 500
},
"eval_heldout": {
"exact_match_K1_pct": 15.151515151515152,
"exact_match_K10_pct": 15.151515151515152,
"token_f1_K1": 0.5692279942279943,
"n": 66
},
"eval_train": {
"exact_match_K1_pct": 51.5625,
"exact_match_K10_pct": 50.520833333333336,
"token_f1_K1": 0.9034226190476198,
"n": 192
},
"decoding_modes": {
"A_full_chain_em_pct": 0.15656565656565657,
"B_pure_noise_1shot_em_pct": 0.11616161616161616,
"C_gold_corrupted_1shot_em_pct": 0.13636363636363635
},
"signal_degradation": {
"raw_xt_tok_acc_at_t2_pct": 100.0,
"model_tok_acc_at_t2_pct": 68.3,
"degradation_ppt": 31.7
},
"state_drift": {
"t2_oracle_tok_acc_pct": 68.3,
"t2_free_tok_acc_pct": 68.3,
"drift_ppt": 0.0,
"oracle_flat_across_noise": true
}
},
"phase4_scheduled": {
"label": "Phase 4 Scheduled",
"architecture": {
"params": null,
"T": 500,
"d_model": 512,
"n_layers": 6
},
"training": {
"steps": 20000,
"lr": 0.0001,
"schedule": "warmup+cosine",
"batch": 64,
"seed": 42,
"T": 500
},
"eval_heldout": {
"exact_match_K1_pct": 13.636363636363637,
"exact_match_K10_pct": 13.636363636363637,
"token_f1_K1": 0.6034992784992784,
"n": 66
},
"eval_train": {
"exact_match_K1_pct": 40.625,
"exact_match_K10_pct": 39.583333333333336,
"token_f1_K1": 0.8827380952380958,
"n": 192
},
"decoding_modes": {
"A_full_chain_em_pct": 0.13636363636363635,
"B_pure_noise_1shot_em_pct": 0.19696969696969696,
"C_gold_corrupted_1shot_em_pct": 0.21717171717171718
},
"signal_degradation": {
"raw_xt_tok_acc_at_t2_pct": 100.0,
"model_tok_acc_at_t2_pct": 76.6,
"degradation_ppt": 23.4
},
"state_drift": {
"t2_oracle_tok_acc_pct": 75.1,
"t2_free_tok_acc_pct": 74.6,
"drift_ppt": 0.51,
"oracle_flat_across_noise": false
}
}
},
"key_findings": [
"Diffusion corruption is NOT the primary bottleneck at near-clean inputs. Raw-x_t achieves 100% token accuracy at t=2; the learned denoiser degrades this to 68-77%.",
"Scaling from 8M to 46.5M parameters reduces held-out exact match from 47% to 15%, despite 40x more training steps.",
"Phase 4 Original has 0.0% state drift (oracle = free-running) and is timestep-agnostic (flat 68.3% accuracy at every noise level).",
"Phase 4 Scheduled has ~0.5 ppt drift and one-shot decoding (B/C = 20-22%) outperforms full chain (A = 14%).",
"K=10 best-of-K never improves EM: only 1.08-1.26 unique candidates out of 10 (pairwise Jaccard > 0.97).",
"The failure is denoiser/model-behavior-limited, not diffusion-corruption-limited."
]
} |