Upload config.json with huggingface_hub

b7d11ec verified 6 days ago

6.19 kB

{
  "model_family": "semsimula-mechanism-ladder",
  "generation": 3,
  "collection": "depth on code",
  "ladder_arm": "none-norc-sr2-l8-k2-code-live",
  "role": "wells at depth: L8-CODE with two V_θ wells instead of eight",
  "variant_of": "semsimula-ladder-live-code-d384-l8-none-norc-sr2",
  "integrator": {
    "name": "baoab_cfc_lowrank",
    "lowrank_damped_flow": true,
    "fixed_gamma": 0.1,
    "first_order_overdamped": false,
    "poisson_modes": 0
  },
  "architecture": "as the arm no-exchange but REVERSE_CHANNEL=False: registers are created and destroyed but never reach the tokens, so every force is the gradient of a scalar potential",
  "params_total": 50164435,
  "context_channels": 5,
  "d": 384,
  "n_layers": 8,
  "dt": 1.0,
  "integration_time_T": 8.0,
  "vocab_size": 50257,
  "block_size": 512,
  "corpus": "codeparrot/codeparrot-clean (Python; the ladder's cache: 450M training tokens, first 2M validation tokens; one tab per four-space indentation level)",
  "total_steps": 19250,
  "schedule_steps": 16250,
  "effective_batch": 32,
  "training_tokens": 315392000,
  "learning_rate": 0.0012,
  "settled_ppl_512": 5.6667,
  "best_ppl_512": 5.56,
  "final_ppl_512": 5.76,
  "best_step": 18000,
  "settled_convention": "exp of the mean log perplexity of the evals at steps 18,000, 18,500 and 19,000",
  "v_theta_wells": 2,
  "so_code_settled_ppl_512": 5.9951,
  "l4_code_settled_ppl_512": 5.4847,
  "l8_code_settled_ppl_512": 5.3545,
  "matched_gpt2_code_settled_ppl_512": 4.96,
  "paired_nats": {
    "spans": [
      "2,000–4,000",
      "8,000–13,000",
      "16,000–19,000"
    ],
    "n": [
      5,
      11,
      7
    ],
    "vs_l8": [
      [
        -0.34,
        0.074
      ],
      [
        0.028,
        0.014
      ],
      [
        0.056,
        0.0015
      ]
    ],
    "vs_l4": [
      [
        0.021,
        0.013
      ],
      [
        0.029,
        0.004
      ],
      [
        0.032,
        0.0016
      ]
    ],
    "vs_so": [
      [
        -0.272,
        0.064
      ],
      [
        -0.07,
        0.009
      ],
      [
        -0.056,
        0.003
      ]
    ]
  },
  "gates_6b7": {
    "g0": 4.79,
    "g1_reset": 17.77,
    "g1_factor": 3.71,
    "l8_g1_factor": 2.72,
    "g2": {
      "1": 96.3,
      "2": 86.98,
      "3": 63.99,
      "4": 39.67,
      "6": 9.44,
      "8": 4.79
    },
    "g3_ppl": {
      "1": 397.13,
      "2": 203.27,
      "3": 146.84,
      "4": 92.45,
      "6": 8.29,
      "8": 4.79,
      "12": 7.63,
      "16": 14.02,
      "24": 41.91,
      "32": 71.65
    },
    "l8_g3": {
      "6": 0.206,
      "12": 0.416,
      "16": 0.783,
      "32": 2.604
    }
  },
  "omega_dt_6b13": {
    "p50": 1.505,
    "layers": [
      3.22,
      1.46,
      1.15,
      1.96,
      1.96,
      1.56,
      0.9,
      0.4
    ],
    "over_wall": 25.5,
    "l8_p50": 1.747,
    "l8_over_wall": 33.8,
    "l8_layer0": 2.08
  },
  "b_k_spectrum_6b4": {
    "pr": 3.03,
    "pr_p05": 2.3,
    "pr_p95": 3.68,
    "ratios": [
      1.0,
      0.754,
      0.599,
      0.465
    ]
  },
  "cell7_ppl": 5.17,
  "flow_r": {
    "rows": [
      [
        "K2-L8-CODE (L = 8, two wells; trained ppl 5.47 on these batches)",
        {
          "k": [
            0.557,
            0.089,
            0.096,
            0.097,
            0.097,
            0.098
          ],
          "stack4": 2.798
        }
      ],
      [
        "L8-CODE (L = 8, eight wells)",
        {
          "k": [
            0.186,
            0.029,
            0.034,
            0.035,
            0.035,
            0.035
          ],
          "stack4": 2.695
        }
      ],
      [
        "L4-CODE (L = 4)",
        {
          "k": [
            0.429,
            0.52,
            0.282,
            0.32,
            0.309,
            0.306
          ],
          "stack4": 2.96
        }
      ],
      [
        "SO-CODE (L = 2)",
        {
          "k": [
            0.97,
            0.526,
            0.3,
            0.384,
            0.376,
            0.369
          ],
          "stack4": 2.858
        }
      ]
    ],
    "kick": [
      0.14,
      0.14,
      0.12,
      0.19,
      0.19,
      0.26,
      0.32,
      0.44
    ],
    "l8_kick": [
      0.15,
      0.14,
      0.08,
      0.13,
      0.15,
      0.2,
      0.26,
      0.39
    ]
  },
  "companion_notes": [
    "Depth_Ladder_and_Matched_Baseline_Protocol.md",
    "Minimum_Well_Count_for_Conservative_Models.md"
  ],
  "pre_registered": "none registered: the arm was set up 2026-11-02 to ask whether depth changes what the six extra wells buy, and launched before calls were written",
  "run_health": "19,250 steps in 19.0 h in one session from step 0 (3.5 s per step with the V_φ-force checkpoint, which recomputes only the pair force in the backward pass, and no layer checkpointing); peak memory 35.0 GB in training, 51.6 GB in evaluation; no non-finite gradient norm at any logged step (L8-CODE had 34 of 45 over steps 1,400–3,600 on the same batches); gradient norm median 0.38 after step 2,000, maximum 2.02 (step 650, warmup); largest rise between logged training losses 0.069 (step 13,050, as in L8-CODE and SO-CODE); 0 watchdog triggers; SCAF leak monitor CLEAN at all 4 audits",
  "cell0": "REVERSE_CHANNEL=False, VPHI_GRAD_PATH='live', XI_GRAD_PATH='live',\nLOWRANK_DAMPED_FLOW=True,                 # SR2\nVPHI_ANALYTIC_FORCE=True, LAYER_CHECKPOINT=False,   # speed switches, same arithmetic\nCORPUS='code', LADDER_STEPS=16_250, EXTRA_FLOOR_STEPS=3_000   # Python code, then 3,000 steps at the floor\nLADDER_L=8                                # eight layers; dt = T / L = 1, T = 8\nV_THETA_WELLS_PER_HEAD=2                  # two wells in the joint V_theta bank instead of eight\nVPHI_FORCE_CHECKPOINT=True                # memory switch, same arithmetic\nSNAPSHOT_STEPS=[1_000, 1_500, 2_000, 2_500, 3_000]   # weights-only snapshots for the well probe",
  "gdrive_folder": "semsimula_fock_cfc_baoab_code_xi5long_topk16_dt32da16_mh4_aniso_dcvt5x2_vtjoint_cgqk_norc_vplive_xilive_sr2_st16250_L8probe_ob_untied_wsd_e5c_plgate_rep0.05_fockreg0.005_g0.1_baoab_cfc_lowrank_lr0p0012_noattn"
}

	{
	"model_family": "semsimula-mechanism-ladder",
	"generation": 3,
	"collection": "depth on code",
	"ladder_arm": "none-norc-sr2-l8-k2-code-live",
	"role": "wells at depth: L8-CODE with two V_θ wells instead of eight",
	"variant_of": "semsimula-ladder-live-code-d384-l8-none-norc-sr2",
	"integrator": {
	"name": "baoab_cfc_lowrank",
	"lowrank_damped_flow": true,
	"fixed_gamma": 0.1,
	"first_order_overdamped": false,
	"poisson_modes": 0
	},
	"architecture": "as the arm no-exchange but REVERSE_CHANNEL=False: registers are created and destroyed but never reach the tokens, so every force is the gradient of a scalar potential",
	"params_total": 50164435,
	"context_channels": 5,
	"d": 384,
	"n_layers": 8,
	"dt": 1.0,
	"integration_time_T": 8.0,
	"vocab_size": 50257,
	"block_size": 512,
	"corpus": "codeparrot/codeparrot-clean (Python; the ladder's cache: 450M training tokens, first 2M validation tokens; one tab per four-space indentation level)",
	"total_steps": 19250,
	"schedule_steps": 16250,
	"effective_batch": 32,
	"training_tokens": 315392000,
	"learning_rate": 0.0012,
	"settled_ppl_512": 5.6667,
	"best_ppl_512": 5.56,
	"final_ppl_512": 5.76,
	"best_step": 18000,
	"settled_convention": "exp of the mean log perplexity of the evals at steps 18,000, 18,500 and 19,000",
	"v_theta_wells": 2,
	"so_code_settled_ppl_512": 5.9951,
	"l4_code_settled_ppl_512": 5.4847,
	"l8_code_settled_ppl_512": 5.3545,
	"matched_gpt2_code_settled_ppl_512": 4.96,
	"paired_nats": {
	"spans": [
	"2,000–4,000",
	"8,000–13,000",
	"16,000–19,000"
	],
	"n": [
	5,
	11,
	7
	],
	"vs_l8": [
	[
	-0.34,
	0.074
	],
	[
	0.028,
	0.014
	],
	[
	0.056,
	0.0015
	]
	],
	"vs_l4": [
	[
	0.021,
	0.013
	],
	[
	0.029,
	0.004
	],
	[
	0.032,
	0.0016
	]
	],
	"vs_so": [
	[
	-0.272,
	0.064
	],
	[
	-0.07,
	0.009
	],
	[
	-0.056,
	0.003
	]
	]
	},
	"gates_6b7": {
	"g0": 4.79,
	"g1_reset": 17.77,
	"g1_factor": 3.71,
	"l8_g1_factor": 2.72,
	"g2": {
	"1": 96.3,
	"2": 86.98,
	"3": 63.99,
	"4": 39.67,
	"6": 9.44,
	"8": 4.79
	},
	"g3_ppl": {
	"1": 397.13,
	"2": 203.27,
	"3": 146.84,
	"4": 92.45,
	"6": 8.29,
	"8": 4.79,
	"12": 7.63,
	"16": 14.02,
	"24": 41.91,
	"32": 71.65
	},
	"l8_g3": {
	"6": 0.206,
	"12": 0.416,
	"16": 0.783,
	"32": 2.604
	}
	},
	"omega_dt_6b13": {
	"p50": 1.505,
	"layers": [
	3.22,
	1.46,
	1.15,
	1.96,
	1.96,
	1.56,
	0.9,
	0.4
	],
	"over_wall": 25.5,
	"l8_p50": 1.747,
	"l8_over_wall": 33.8,
	"l8_layer0": 2.08
	},
	"b_k_spectrum_6b4": {
	"pr": 3.03,
	"pr_p05": 2.3,
	"pr_p95": 3.68,
	"ratios": [
	1.0,
	0.754,
	0.599,
	0.465
	]
	},
	"cell7_ppl": 5.17,
	"flow_r": {
	"rows": [
	[
	"K2-L8-CODE (L = 8, two wells; trained ppl 5.47 on these batches)",
	{
	"k": [
	0.557,
	0.089,
	0.096,
	0.097,
	0.097,
	0.098
	],
	"stack4": 2.798
	}
	],
	[
	"L8-CODE (L = 8, eight wells)",
	{
	"k": [
	0.186,
	0.029,
	0.034,
	0.035,
	0.035,
	0.035
	],
	"stack4": 2.695
	}
	],
	[
	"L4-CODE (L = 4)",
	{
	"k": [
	0.429,
	0.52,
	0.282,
	0.32,
	0.309,
	0.306
	],
	"stack4": 2.96
	}
	],
	[
	"SO-CODE (L = 2)",
	{
	"k": [
	0.97,
	0.526,
	0.3,
	0.384,
	0.376,
	0.369
	],
	"stack4": 2.858
	}
	]
	],
	"kick": [
	0.14,
	0.14,
	0.12,
	0.19,
	0.19,
	0.26,
	0.32,
	0.44
	],
	"l8_kick": [
	0.15,
	0.14,
	0.08,
	0.13,
	0.15,
	0.2,
	0.26,
	0.39
	]
	},
	"companion_notes": [
	"Depth_Ladder_and_Matched_Baseline_Protocol.md",
	"Minimum_Well_Count_for_Conservative_Models.md"
	],
	"pre_registered": "none registered: the arm was set up 2026-11-02 to ask whether depth changes what the six extra wells buy, and launched before calls were written",
	"run_health": "19,250 steps in 19.0 h in one session from step 0 (3.5 s per step with the V_φ-force checkpoint, which recomputes only the pair force in the backward pass, and no layer checkpointing); peak memory 35.0 GB in training, 51.6 GB in evaluation; no non-finite gradient norm at any logged step (L8-CODE had 34 of 45 over steps 1,400–3,600 on the same batches); gradient norm median 0.38 after step 2,000, maximum 2.02 (step 650, warmup); largest rise between logged training losses 0.069 (step 13,050, as in L8-CODE and SO-CODE); 0 watchdog triggers; SCAF leak monitor CLEAN at all 4 audits",
	"cell0": "REVERSE_CHANNEL=False, VPHI_GRAD_PATH='live', XI_GRAD_PATH='live',\nLOWRANK_DAMPED_FLOW=True, # SR2\nVPHI_ANALYTIC_FORCE=True, LAYER_CHECKPOINT=False, # speed switches, same arithmetic\nCORPUS='code', LADDER_STEPS=16_250, EXTRA_FLOOR_STEPS=3_000 # Python code, then 3,000 steps at the floor\nLADDER_L=8 # eight layers; dt = T / L = 1, T = 8\nV_THETA_WELLS_PER_HEAD=2 # two wells in the joint V_theta bank instead of eight\nVPHI_FORCE_CHECKPOINT=True # memory switch, same arithmetic\nSNAPSHOT_STEPS=[1_000, 1_500, 2_000, 2_500, 3_000] # weights-only snapshots for the well probe",
	"gdrive_folder": "semsimula_fock_cfc_baoab_code_xi5long_topk16_dt32da16_mh4_aniso_dcvt5x2_vtjoint_cgqk_norc_vplive_xilive_sr2_st16250_L8probe_ob_untied_wsd_e5c_plgate_rep0.05_fockreg0.005_g0.1_baoab_cfc_lowrank_lr0p0012_noattn"
	}