| { |
| "E3A_bind_s1": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "E3A_bind_s1", |
| "batch": 128, |
| "cfg": { |
| "T_ev": 48, |
| "d": 192, |
| "enc_layers": 3, |
| "gate_bias": 0.0, |
| "heads": 6, |
| "mode": "bind", |
| "n_slots": 192, |
| "vocab": 8192 |
| }, |
| "group": "experiment2_stageA", |
| "kind": "bind", |
| "lr": 0.0003, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 4385515, |
| "role": "Stage A, binding-local writes, seed 1", |
| "run": "E3A_bind_s1", |
| "seed": 1, |
| "source_run_directory": "E3A_bind_s1", |
| "state_dict_tensor_total": 4385515, |
| "train_args": { |
| "batch": 128, |
| "eval_lifetimes": 400, |
| "lr": 0.0003, |
| "mode": "bind", |
| "run": "E3A_bind_s1", |
| "seed": 1, |
| "updates": 2600 |
| }, |
| "train_corpus": "e3_train", |
| "updates": 2600 |
| }, |
| "E3A_bind_s2": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "E3A_bind_s2", |
| "batch": 128, |
| "cfg": { |
| "T_ev": 48, |
| "d": 192, |
| "enc_layers": 3, |
| "gate_bias": 0.0, |
| "heads": 6, |
| "mode": "bind", |
| "n_slots": 192, |
| "vocab": 8192 |
| }, |
| "group": "experiment2_stageA", |
| "kind": "bind", |
| "lr": 0.0003, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 4385515, |
| "role": "Stage A, binding-local writes, seed 2", |
| "run": "E3A_bind_s2", |
| "seed": 2, |
| "source_run_directory": "E3A_bind_s2", |
| "state_dict_tensor_total": 4385515, |
| "train_args": { |
| "batch": 128, |
| "eval_lifetimes": 400, |
| "lr": 0.0003, |
| "mode": "bind", |
| "run": "E3A_bind_s2", |
| "seed": 2, |
| "updates": 2600 |
| }, |
| "train_corpus": "e3_train", |
| "updates": 2600 |
| }, |
| "E3A_bind_s3": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "E3A_bind_s3", |
| "batch": 128, |
| "cfg": { |
| "T_ev": 48, |
| "d": 192, |
| "enc_layers": 3, |
| "gate_bias": 0.0, |
| "heads": 6, |
| "mode": "bind", |
| "n_slots": 192, |
| "vocab": 8192 |
| }, |
| "group": "experiment2_stageA", |
| "kind": "bind", |
| "lr": 0.0003, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 4385515, |
| "role": "Stage A, binding-local writes, seed 3", |
| "run": "E3A_bind_s3", |
| "seed": 3, |
| "source_run_directory": "E3A_bind_s3", |
| "state_dict_tensor_total": 4385515, |
| "train_args": { |
| "batch": 128, |
| "eval_lifetimes": 400, |
| "lr": 0.0003, |
| "mode": "bind", |
| "run": "E3A_bind_s3", |
| "seed": 3, |
| "updates": 2600 |
| }, |
| "train_corpus": "e3_train", |
| "updates": 2600 |
| }, |
| "E3A_unbound_s1": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "E3A_unbound_s1", |
| "batch": 128, |
| "cfg": { |
| "T_ev": 48, |
| "d": 192, |
| "enc_layers": 3, |
| "gate_bias": 0.0, |
| "heads": 6, |
| "mode": "unbound", |
| "n_slots": 192, |
| "vocab": 8192 |
| }, |
| "group": "experiment2_stageA", |
| "kind": "unbound", |
| "lr": 0.0003, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 4385515, |
| "role": "Stage A, capacity-matched unstructured control, seed 1", |
| "run": "E3A_unbound_s1", |
| "seed": 1, |
| "source_run_directory": "E3A_unbound_s1", |
| "state_dict_tensor_total": 4385515, |
| "train_args": { |
| "batch": 128, |
| "eval_lifetimes": 400, |
| "lr": 0.0003, |
| "mode": "unbound", |
| "run": "E3A_unbound_s1", |
| "seed": 1, |
| "updates": 2600 |
| }, |
| "train_corpus": "e3_train", |
| "updates": 2600 |
| }, |
| "E3A_unbound_s2": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "E3A_unbound_s2", |
| "batch": 128, |
| "cfg": { |
| "T_ev": 48, |
| "d": 192, |
| "enc_layers": 3, |
| "gate_bias": 0.0, |
| "heads": 6, |
| "mode": "unbound", |
| "n_slots": 192, |
| "vocab": 8192 |
| }, |
| "group": "experiment2_stageA", |
| "kind": "unbound", |
| "lr": 0.0003, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 4385515, |
| "role": "Stage A, capacity-matched unstructured control, seed 2", |
| "run": "E3A_unbound_s2", |
| "seed": 2, |
| "source_run_directory": "E3A_unbound_s2", |
| "state_dict_tensor_total": 4385515, |
| "train_args": { |
| "batch": 128, |
| "eval_lifetimes": 400, |
| "lr": 0.0003, |
| "mode": "unbound", |
| "run": "E3A_unbound_s2", |
| "seed": 2, |
| "updates": 2600 |
| }, |
| "train_corpus": "e3_train", |
| "updates": 2600 |
| }, |
| "E3A_unbound_s3": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "E3A_unbound_s3", |
| "batch": 128, |
| "cfg": { |
| "T_ev": 48, |
| "d": 192, |
| "enc_layers": 3, |
| "gate_bias": 0.0, |
| "heads": 6, |
| "mode": "unbound", |
| "n_slots": 192, |
| "vocab": 8192 |
| }, |
| "group": "experiment2_stageA", |
| "kind": "unbound", |
| "lr": 0.0003, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 4385515, |
| "role": "Stage A, capacity-matched unstructured control, seed 3", |
| "run": "E3A_unbound_s3", |
| "seed": 3, |
| "source_run_directory": "E3A_unbound_s3", |
| "state_dict_tensor_total": 4385515, |
| "train_args": { |
| "batch": 128, |
| "eval_lifetimes": 400, |
| "lr": 0.0003, |
| "mode": "unbound", |
| "run": "E3A_unbound_s3", |
| "seed": 3, |
| "updates": 2600 |
| }, |
| "train_corpus": "e3_train", |
| "updates": 2600 |
| }, |
| "E3B_bind_s1": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "E3B_bind_s1", |
| "batch": 96, |
| "cfg": { |
| "T_ev": 48, |
| "d": 384, |
| "enc_layers": 8, |
| "full_heads": true, |
| "gate_bias": 0.0, |
| "heads": 8, |
| "mode": "bind", |
| "n_slots": 192, |
| "vocab": 8192 |
| }, |
| "group": "experiment2_stageB", |
| "kind": "bind", |
| "lr": 0.0004, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 24265440, |
| "role": "Stage B, binding-local writes, seed 1", |
| "run": "E3B_bind_s1", |
| "seed": 1, |
| "source_run_directory": "E3B_bind_s1", |
| "state_dict_tensor_total": 27411168, |
| "train_args": { |
| "batch": 96, |
| "lr": 0.0004, |
| "mode": "bind", |
| "run": "E3B_bind_s1", |
| "seed": 1, |
| "updates": 6000 |
| }, |
| "train_corpus": "e3_train", |
| "updates": 6000 |
| }, |
| "E3B_bind_s2": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "E3B_bind_s2", |
| "batch": 96, |
| "cfg": { |
| "T_ev": 48, |
| "d": 384, |
| "enc_layers": 8, |
| "full_heads": true, |
| "gate_bias": 0.0, |
| "heads": 8, |
| "mode": "bind", |
| "n_slots": 192, |
| "vocab": 8192 |
| }, |
| "group": "experiment2_stageB", |
| "kind": "bind", |
| "lr": 0.0004, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 24265440, |
| "role": "Stage B, binding-local writes, seed 2", |
| "run": "E3B_bind_s2", |
| "seed": 2, |
| "source_run_directory": "E3B_bind_s2", |
| "state_dict_tensor_total": 27411168, |
| "train_args": { |
| "batch": 96, |
| "lr": 0.0004, |
| "mode": "bind", |
| "run": "E3B_bind_s2", |
| "seed": 2, |
| "updates": 6000 |
| }, |
| "train_corpus": "e3_train", |
| "updates": 6000 |
| }, |
| "E3B_bind_s3": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "E3B_bind_s3", |
| "batch": 96, |
| "cfg": { |
| "T_ev": 48, |
| "d": 384, |
| "enc_layers": 8, |
| "full_heads": true, |
| "gate_bias": 0.0, |
| "heads": 8, |
| "mode": "bind", |
| "n_slots": 192, |
| "vocab": 8192 |
| }, |
| "group": "experiment2_stageB", |
| "kind": "bind", |
| "lr": 0.0004, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 24265440, |
| "role": "Stage B, binding-local writes, seed 3", |
| "run": "E3B_bind_s3", |
| "seed": 3, |
| "source_run_directory": "E3B_bind_s3", |
| "state_dict_tensor_total": 27411168, |
| "train_args": { |
| "batch": 96, |
| "lr": 0.0004, |
| "mode": "bind", |
| "run": "E3B_bind_s3", |
| "seed": 3, |
| "updates": 6000 |
| }, |
| "train_corpus": "e3_train", |
| "updates": 6000 |
| }, |
| "E3B_unbound_s1": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "E3B_unbound_s1", |
| "batch": 96, |
| "cfg": { |
| "T_ev": 48, |
| "d": 384, |
| "enc_layers": 8, |
| "full_heads": true, |
| "gate_bias": 0.0, |
| "heads": 8, |
| "mode": "unbound", |
| "n_slots": 192, |
| "vocab": 8192 |
| }, |
| "group": "experiment2_stageB", |
| "kind": "unbound", |
| "lr": 0.0004, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 24265440, |
| "role": "Stage B, capacity/parameter-matched unstructured control, seed 1", |
| "run": "E3B_unbound_s1", |
| "seed": 1, |
| "source_run_directory": "E3B_unbound_s1", |
| "state_dict_tensor_total": 27411168, |
| "train_args": { |
| "batch": 96, |
| "lr": 0.0004, |
| "mode": "unbound", |
| "run": "E3B_unbound_s1", |
| "seed": 1, |
| "updates": 6000 |
| }, |
| "train_corpus": "e3_train", |
| "updates": 6000 |
| }, |
| "E3B_unbound_s2": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "E3B_unbound_s2", |
| "batch": 96, |
| "cfg": { |
| "T_ev": 48, |
| "d": 384, |
| "enc_layers": 8, |
| "full_heads": true, |
| "gate_bias": 0.0, |
| "heads": 8, |
| "mode": "unbound", |
| "n_slots": 192, |
| "vocab": 8192 |
| }, |
| "group": "experiment2_stageB", |
| "kind": "unbound", |
| "lr": 0.0004, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 24265440, |
| "role": "Stage B, capacity/parameter-matched unstructured control, seed 2", |
| "run": "E3B_unbound_s2", |
| "seed": 2, |
| "source_run_directory": "E3B_unbound_s2", |
| "state_dict_tensor_total": 27411168, |
| "train_args": { |
| "batch": 96, |
| "lr": 0.0004, |
| "mode": "unbound", |
| "run": "E3B_unbound_s2", |
| "seed": 2, |
| "updates": 6000 |
| }, |
| "train_corpus": "e3_train", |
| "updates": 6000 |
| }, |
| "E3B_unbound_s3": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "E3B_unbound_s3", |
| "batch": 96, |
| "cfg": { |
| "T_ev": 48, |
| "d": 384, |
| "enc_layers": 8, |
| "full_heads": true, |
| "gate_bias": 0.0, |
| "heads": 8, |
| "mode": "unbound", |
| "n_slots": 192, |
| "vocab": 8192 |
| }, |
| "group": "experiment2_stageB", |
| "kind": "unbound", |
| "lr": 0.0004, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 24265440, |
| "role": "Stage B, capacity/parameter-matched unstructured control, seed 3", |
| "run": "E3B_unbound_s3", |
| "seed": 3, |
| "source_run_directory": "E3B_unbound_s3", |
| "state_dict_tensor_total": 27411168, |
| "train_args": { |
| "batch": 96, |
| "lr": 0.0004, |
| "mode": "unbound", |
| "run": "E3B_unbound_s3", |
| "seed": 3, |
| "updates": 6000 |
| }, |
| "train_corpus": "e3_train", |
| "updates": 6000 |
| }, |
| "G_CONVEX_FREE": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "G_CONVEX_FREE", |
| "batch": 160, |
| "cfg": { |
| "K": 4, |
| "T_ev": 48, |
| "d": 384, |
| "enc_layers": 7, |
| "heads": 8, |
| "n_workspace": 4, |
| "state_tau": 16.0, |
| "update_rule": "convex", |
| "vocab": 8192, |
| "write_mode": "independent", |
| "write_temp": 1.0 |
| }, |
| "group": "experiment1", |
| "kind": "pnsr", |
| "lr": 0.0004, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 24614453, |
| "role": "E2-GATED arm: convex gated update, no clamp", |
| "run": "G_CONVEX_FREE", |
| "seed": 1, |
| "source_run_directory": "G_CONVEX_FREE", |
| "state_dict_tensor_total": 27760181, |
| "train_args": { |
| "batch": 160, |
| "gate_bias": 0.0, |
| "hold_frac": null, |
| "lr": 0.0004, |
| "model": "pnsr", |
| "resume_from": null, |
| "run": "G_CONVEX_FREE", |
| "seed": 1, |
| "state_tau": 16.0, |
| "update_rule": "convex", |
| "updates": 6000, |
| "w_enum": 2.0, |
| "w_other": 1.0 |
| }, |
| "train_corpus": "train", |
| "updates": 6000 |
| }, |
| "G_CONVEX_RETAIN": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "G_CONVEX_RETAIN", |
| "batch": 160, |
| "cfg": { |
| "K": 4, |
| "T_ev": 48, |
| "d": 384, |
| "enc_layers": 7, |
| "heads": 8, |
| "n_workspace": 4, |
| "state_tau": 16.0, |
| "update_rule": "convex", |
| "vocab": 8192, |
| "write_mode": "independent", |
| "write_temp": 1.0 |
| }, |
| "group": "experiment1", |
| "kind": "pnsr", |
| "lr": 0.0004, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 24614453, |
| "role": "E2-GATED arm: convex gated update, retention-biased", |
| "run": "G_CONVEX_RETAIN", |
| "seed": 1, |
| "source_run_directory": "G_CONVEX_RETAIN", |
| "state_dict_tensor_total": 27760181, |
| "train_args": { |
| "batch": 160, |
| "gate_bias": -2.0, |
| "hold_frac": null, |
| "lr": 0.0004, |
| "model": "pnsr", |
| "resume_from": null, |
| "run": "G_CONVEX_RETAIN", |
| "seed": 1, |
| "update_rule": "convex", |
| "updates": 6000, |
| "w_enum": 2.0, |
| "w_other": 1.0 |
| }, |
| "train_corpus": "train", |
| "updates": 6000 |
| }, |
| "G_TAU512": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "G_TAU512", |
| "batch": 160, |
| "cfg": { |
| "K": 4, |
| "T_ev": 48, |
| "d": 384, |
| "enc_layers": 7, |
| "heads": 8, |
| "n_workspace": 4, |
| "state_tau": 512.0, |
| "update_rule": "additive_clamp", |
| "vocab": 8192, |
| "write_mode": "independent", |
| "write_temp": 1.0 |
| }, |
| "group": "experiment1", |
| "kind": "pnsr", |
| "lr": 0.0004, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 24614453, |
| "role": "isolating control: same additive rule, clamp raised 16 -> 512", |
| "run": "G_TAU512", |
| "seed": 1, |
| "source_run_directory": "G_TAU512", |
| "state_dict_tensor_total": 27760181, |
| "train_args": { |
| "batch": 160, |
| "gate_bias": 0.0, |
| "hold_frac": null, |
| "lr": 0.0004, |
| "model": "pnsr", |
| "resume_from": null, |
| "run": "G_TAU512", |
| "seed": 1, |
| "state_tau": 512.0, |
| "update_rule": "additive_clamp", |
| "updates": 6000, |
| "w_enum": 2.0, |
| "w_other": 1.0 |
| }, |
| "train_corpus": "train", |
| "updates": 6000 |
| }, |
| "PNSR_K1_fin": { |
| "admission_note": "amendment A6: held at stage 1 for a protocol-matched comparison with PNSR_K4_fin", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "PNSR_K1_v2", |
| "batch": 160, |
| "cfg": { |
| "K": 1, |
| "T_ev": 48, |
| "d": 384, |
| "enc_layers": 7, |
| "heads": 8, |
| "n_workspace": 4, |
| "state_tau": 16.0, |
| "vocab": 8192 |
| }, |
| "group": "experiment1", |
| "kind": "pnsr_k1", |
| "lr": 0.0004, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 24614453, |
| "role": "deliberation-depth control K=1 (E5)", |
| "run": "PNSR_K1_fin", |
| "seed": 1, |
| "source_run_directory": "PNSR_K1_fin", |
| "state_dict_tensor_total": 27760181, |
| "train_args": { |
| "batch": 160, |
| "lr": 0.0004, |
| "model": "pnsr_k1", |
| "run": "PNSR_K1_v2", |
| "seed": 1, |
| "updates": 6000 |
| }, |
| "train_corpus": "train", |
| "updates": 6000 |
| }, |
| "PNSR_K4_fin": { |
| "admission_note": "amendment A6: the 8x-enum stage-2 continuation failed its own preregistered admission rule, so the stage-1 checkpoint is the admitted one", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "PNSR_K4_v2", |
| "batch": 160, |
| "cfg": { |
| "K": 4, |
| "T_ev": 48, |
| "d": 384, |
| "enc_layers": 7, |
| "heads": 8, |
| "n_workspace": 4, |
| "state_tau": 16.0, |
| "vocab": 8192 |
| }, |
| "group": "experiment1", |
| "kind": "pnsr", |
| "lr": 0.0004, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 24614453, |
| "role": "PNSR typed learned state, seed 1 (additive write, clamp tau=16)", |
| "run": "PNSR_K4_fin", |
| "seed": 1, |
| "source_run_directory": "PNSR_K4_fin", |
| "state_dict_tensor_total": 27760181, |
| "train_args": { |
| "batch": 160, |
| "lr": 0.0004, |
| "model": "pnsr", |
| "run": "PNSR_K4_v2", |
| "seed": 1, |
| "updates": 6000 |
| }, |
| "train_corpus": "train", |
| "updates": 6000 |
| }, |
| "PNSR_K4_s2": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "PNSR_K4_s2", |
| "batch": 160, |
| "cfg": { |
| "K": 4, |
| "T_ev": 48, |
| "d": 384, |
| "enc_layers": 7, |
| "heads": 8, |
| "n_workspace": 4, |
| "state_tau": 16.0, |
| "update_rule": "additive_clamp", |
| "vocab": 8192, |
| "write_mode": "independent", |
| "write_temp": 1.0 |
| }, |
| "group": "experiment1", |
| "kind": "pnsr", |
| "lr": 0.0004, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 24614453, |
| "role": "PNSR typed learned state, seed 2", |
| "run": "PNSR_K4_s2", |
| "seed": 2, |
| "source_run_directory": "PNSR_K4_s2", |
| "state_dict_tensor_total": 27760181, |
| "train_args": { |
| "batch": 160, |
| "gate_bias": 0.0, |
| "hold_frac": null, |
| "lr": 0.0004, |
| "model": "pnsr", |
| "resume_from": null, |
| "run": "PNSR_K4_s2", |
| "seed": 2, |
| "state_tau": 16.0, |
| "update_rule": "additive_clamp", |
| "updates": 6000, |
| "w_enum": 2.0, |
| "w_other": 1.0 |
| }, |
| "train_corpus": "train", |
| "updates": 6000 |
| }, |
| "PNSR_K4_s3": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "PNSR_K4_s3", |
| "batch": 160, |
| "cfg": { |
| "K": 4, |
| "T_ev": 48, |
| "d": 384, |
| "enc_layers": 7, |
| "heads": 8, |
| "n_workspace": 4, |
| "state_tau": 16.0, |
| "update_rule": "additive_clamp", |
| "vocab": 8192, |
| "write_mode": "independent", |
| "write_temp": 1.0 |
| }, |
| "group": "experiment1", |
| "kind": "pnsr", |
| "lr": 0.0004, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 24614453, |
| "role": "PNSR typed learned state, seed 3", |
| "run": "PNSR_K4_s3", |
| "seed": 3, |
| "source_run_directory": "PNSR_K4_s3", |
| "state_dict_tensor_total": 27760181, |
| "train_args": { |
| "batch": 160, |
| "gate_bias": 0.0, |
| "hold_frac": null, |
| "lr": 0.0004, |
| "model": "pnsr", |
| "resume_from": null, |
| "run": "PNSR_K4_s3", |
| "seed": 3, |
| "state_tau": 16.0, |
| "update_rule": "additive_clamp", |
| "updates": 6000, |
| "w_enum": 2.0, |
| "w_other": 1.0 |
| }, |
| "train_corpus": "train", |
| "updates": 6000 |
| }, |
| "PNSR_ON_E3": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "PNSR_ON_E3", |
| "batch": 96, |
| "cfg": { |
| "K": 4, |
| "T_ev": 48, |
| "d": 384, |
| "enc_layers": 7, |
| "heads": 8, |
| "n_workspace": 4, |
| "state_tau": 16.0, |
| "update_rule": "additive_clamp", |
| "vocab": 8192, |
| "write_mode": "independent", |
| "write_temp": 1.0 |
| }, |
| "group": "corrected_corpus", |
| "kind": "pnsr", |
| "lr": 0.0004, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 24614453, |
| "role": "E1 architecture retrained from scratch on the corrected corpus", |
| "run": "PNSR_ON_E3", |
| "seed": 1, |
| "source_run_directory": "PNSR_ON_E3", |
| "state_dict_tensor_total": 27760181, |
| "train_args": { |
| "batch": 96, |
| "gate_bias": 0.0, |
| "hold_frac": null, |
| "lr": 0.0004, |
| "model": "pnsr", |
| "resume_from": null, |
| "run": "PNSR_ON_E3", |
| "seed": 1, |
| "shards_split": "e3_train", |
| "state_tau": 16.0, |
| "update_rule": "additive_clamp", |
| "updates": 6000, |
| "w_enum": 2.0, |
| "w_other": 1.0 |
| }, |
| "train_corpus": "e3_train", |
| "updates": 6000 |
| }, |
| "RMT_s1": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "RMT_s1", |
| "batch": 160, |
| "cfg": { |
| "T_ev": 48, |
| "d": 384, |
| "heads": 8, |
| "layers": 10, |
| "n_mem": 32, |
| "vocab": 8192 |
| }, |
| "group": "experiment1", |
| "kind": "rmt", |
| "lr": 0.0003, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 23945525, |
| "role": "memory-token recurrent baseline (RMT-style), seed 1", |
| "run": "RMT_s1", |
| "seed": 1, |
| "source_run_directory": "RMT_s1", |
| "state_dict_tensor_total": 27091253, |
| "train_args": { |
| "batch": 160, |
| "gate_bias": 0.0, |
| "hold_frac": null, |
| "lr": 0.0003, |
| "model": "rmt", |
| "resume_from": null, |
| "run": "RMT_s1", |
| "seed": 1, |
| "state_tau": 16.0, |
| "update_rule": "additive_clamp", |
| "updates": 6000, |
| "w_enum": 2.0, |
| "w_other": 1.0 |
| }, |
| "train_corpus": "train", |
| "updates": 6000 |
| }, |
| "TX768_fin": { |
| "admission_note": "amendment A7: v2/v3 invalid (v3 was frozen mid-run by the A3 skip-guard, AUDIT 2); v4 retrained with the guard removed, lr 1.5e-4, exposure-matched; best.pt (update 18250) is the admitted checkpoint", |
| "admitted_checkpoint": "best.pt", |
| "admitted_from_run": "TX768_v4", |
| "batch": 384, |
| "cfg": { |
| "d": 384, |
| "heads": 8, |
| "layers": 10, |
| "vocab": 8192, |
| "window": 768 |
| }, |
| "group": "experiment1", |
| "kind": "tx768", |
| "lr": 0.00015, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 23892917, |
| "role": "transcript-window baseline, 768 tokens (~40 events)", |
| "run": "TX768_fin", |
| "seed": 1, |
| "source_run_directory": "TX768_fin", |
| "state_dict_tensor_total": 27038645, |
| "train_args": { |
| "batch": 384, |
| "hold_frac": 0.0, |
| "lr": 0.00015, |
| "model": "tx768", |
| "resume_from": null, |
| "run": "TX768_v4", |
| "seed": 1, |
| "updates": 18850, |
| "w_enum": 2.0, |
| "w_other": 1.0 |
| }, |
| "train_corpus": "train", |
| "updates": 18850 |
| }, |
| "TXE_ON_E3": { |
| "admission_note": "the run's own final checkpoint", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "TXE_ON_E3", |
| "batch": 384, |
| "cfg": { |
| "d": 384, |
| "heads": 8, |
| "layers": 10, |
| "vocab": 8192, |
| "window": 64 |
| }, |
| "group": "corrected_corpus", |
| "kind": "txe", |
| "lr": 0.0003, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 23622581, |
| "role": "memory-free control on the corrected corpus (Experiment 2 comparator)", |
| "run": "TXE_ON_E3", |
| "seed": 1, |
| "source_run_directory": "TXE_ON_E3", |
| "state_dict_tensor_total": 26768309, |
| "train_args": { |
| "batch": 384, |
| "gate_bias": 0.0, |
| "hold_frac": null, |
| "lr": 0.0003, |
| "model": "txe", |
| "resume_from": null, |
| "run": "TXE_ON_E3", |
| "seed": 1, |
| "shards_split": "e3_train", |
| "state_tau": 16.0, |
| "update_rule": "additive_clamp", |
| "updates": 20000, |
| "w_enum": 2.0, |
| "w_other": 1.0 |
| }, |
| "train_corpus": "e3_train", |
| "updates": 20000 |
| }, |
| "TXE_fin": { |
| "admission_note": "TXE_v2c diverged and was killed before completion (amendment A5 auto-fail)", |
| "admitted_checkpoint": "final.pt", |
| "admitted_from_run": "TXE_v2", |
| "batch": 384, |
| "cfg": { |
| "d": 384, |
| "heads": 8, |
| "layers": 10, |
| "vocab": 8192, |
| "window": 64 |
| }, |
| "group": "experiment1", |
| "kind": "txe", |
| "lr": 0.0003, |
| "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.", |
| "params": 23622581, |
| "role": "memory-free control: current event only", |
| "run": "TXE_fin", |
| "seed": 1, |
| "source_run_directory": "TXE_fin", |
| "state_dict_tensor_total": 26768309, |
| "train_args": { |
| "batch": 384, |
| "lr": 0.0003, |
| "model": "txe", |
| "run": "TXE_v2", |
| "seed": 1, |
| "updates": 20000 |
| }, |
| "train_corpus": "train", |
| "updates": 20000 |
| } |
| } |
|
|
