{ "model_family": "semsimula-mechanism-ladder", "generation": 3, "collection": "first vs second order", "ladder_arm": "none-norc-sr2-fob-code-live", "pair": "code", "role": "first order (FO-b), no wells, Python code", "pair_partner": "semsimula-ladder-live-code-d384-l2-none-norc-sr2", "variant_of": "semsimula-ladder-live-owt-d384-l2-none-norc-sr2", "integrator": { "name": "baoab_cfc_lowrank", "lowrank_damped_flow": true, "fixed_gamma": 0.1, "first_order_overdamped": true, "first_order_memoryless": 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": 76572109, "d": 384, "n_layers": 2, "dt": 4.0, "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": 6.5487, "best_ppl_512": 6.43, "final_ppl_512": 6.65, "best_step": 18000, "settled_convention": "mean of the evals at steps 18,000, 18,500 and 19,000", "pair_partner_settled_ppl_512": 5.9951, "matched_gpt2_code_settled_ppl_512": 4.96, "paired_first_vs_second_order": { "stable_end": 11.1, "decay": 10.5, "floor": 9.2, "floor_range": [ 9.1, 9.5 ], "floor_sd_nats": 0.0015, "floor_trend_nats": -0.003, "corr_floor": 0.996, "windows_worse_pct": 98.2, "settling": { "so": [ -0.02, 0.015 ], "fo": [ -0.023, 0.015 ] } }, "copy_split": { "ratio": 1.09, "all": 0.086, "rows": [ [ "identifier-like (names and keywords)", 49.3, 0.152, 86.6, 0.253 ], [ "indentation, newlines, spaces", 21.9, 0.008, 2.0, 0.042 ], [ "numbers, punctuation and other", 28.8, 0.034, 11.3, 0.182 ], [ "first occurrence in the window", 27.7, 0.085, 27.3, 0.189 ], [ "earlier occurrence 1–16 tokens back", 34.8, 0.043, 17.1, 0.109 ], [ "earlier occurrence 17–64 tokens back", 25.2, 0.102, 29.8, 0.208 ], [ "earlier occurrence 65–256 tokens back", 10.8, 0.178, 22.3, 0.342 ], [ "earlier occurrence 257–512 tokens back", 1.5, 0.197, 3.5, 0.395 ] ], "fixed_pos": [ 0.042, 0.102, 0.176, 0.197 ], "xi_far": { "fob": [ 0.245, 28.8, 25.7 ], "so": [ 0.223, 26.3, 36.1 ] } }, "omega_dt_6b13": { "p50": 0.86, "layers": [ 1.17, 0.002 ], "over_wall": 22.9 }, "pre_registered": "CODE1 (FO-b-CODE settles more than 2% above SO-CODE), called 65%: HIT, +9.2%. CODE2 (the gap is larger than OpenWebText's +4.9%), called 35%: HIT. Scored without an extension: both arms fail the settling check by its letter (−0.020 and −0.023 nats over the last 3,000 steps, against a residual scatter of 0.015) but fall together, and the paired gap moves by 0.003 nats over that window. CODE3–CODE5 were scored on the two-well pair on code (all three HIT; see PM2-CODE and FO-b-PM2-CODE)", "run_health": "19,250 steps in 5.65 h in one session; 0 watchdog triggers; largest rise between logged training losses 0.073, at the same step as SO-CODE's (the batches are paired); gradient norm median 0.67, maximum 1.83 (in warmup); SCAF leak monitor CLEAN at all 4 audits; trained leak probe CLEAN at step 10,000", "gates_6b7": { "g1": 0, "g2_n3": 3.6, "g2_n8": 3987, "g3_n3": 0.277 }, "cell0": "REVERSE_CHANNEL=False, VPHI_GRAD_PATH='live', XI_GRAD_PATH='live',\nLOWRANK_DAMPED_FLOW=True, # SR2\nFIRST_ORDER_OVERDAMPED=True, # tag 'fob'\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", "gdrive_folder": "semsimula_fock_cfc_baoab_code_xi5long_topk16_dt32da16_mh4_aniso_dcvt5x8_vtjoint_cgqk_norc_vplive_xilive_sr2_fob_st16250_L2probe_ob_untied_wsd_e5c_plgate_rep0.05_fockreg0.005_g0.1_baoab_cfc_lowrank_idt4_lr0p0012_noattn" }