Text Generation
PyTorch
English
research
semsimula
conservative-language-model
scalar-potential
lagrangian-mechanics
energy-based-model
physics-informed
openwebtext
mechanism-ablation
ladder-study
pre-registered
live-gradients
generation-3
poisson-mode-registers
exact-integrator
matched-token-budget
fock-space
parflm
anisotropic-gaussian-potential
cfc
baoab
fully-conservative
riemannian-geometry
riemannian-geodesics
jacobi-metric
non-transformer
attention-free
constant-memory-inference
Eval Results (legacy)
| model: ...cgqk_norc_vplive_xilive_pm64_pmx16_sr2_L2probe_ob_untied_wsd_e5c_plgate_rep0.05_fockreg0.005_g0.1_baoab_cfc_lowrank_idt4_lr0p0012_noattn | |
| 1. FUTURE PERTURBATION (max |d logit| at positions <= p; must be exactly 0) | |
| cut p= 31: max |d| over 3 draws = 0.000e+00 | |
| cut p=127: max |d| over 3 draws = 0.000e+00 | |
| cut p=255: max |d| over 3 draws = 0.000e+00 | |
| cut p=383: max |d| over 3 draws = 0.000e+00 | |
| cut p=500: max |d| over 3 draws = 0.000e+00 | |
| -> CLEAN (worst 0.000e+00) | |
| 2. BATCH INDEPENDENCE (replace sequence 2 entirely; sequences 0,1 must not move) | |
| max |d| on the untouched sequences = 0.000e+00 -> CLEAN | |
| 3. PREFIX-ONLY SCORING (feed x[:t+1] only; compare the logit at t) | |
| positions 0..511 (8 sampled x 3 seqs): max |d logit| = 5.999e-03, leak tax = -1.00e-05 nats -> CLEAN | |
| 4. VALIDATION PPL on the local cache (load check; 12 batches x 4 x 512) | |
| val loss 3.9901 PPL 54.06 | |