Hippius
maincontainerPublic

tired-cascade/my_gen_v7

sha256:81dbb0cf45acedd8a1dab50f960b10bdc42e6b93bf4b817ba3b70d39b8b9e227·Indexed 26d ago

Layers

7

Total size

52.3 KB

Files

7

Quantization

README.md

869 B

my_gen_v7

Fork of my_gen_v6 with k01, k09, and k13 removed from top-level sampling. Their combined weight, 0.041108688881968236, is split equally across the three prioritized families:

  • k14: 0.435864216754905
  • k15: 0.1338804110868888
  • k21: 0.023702896293989413

The unchanged active families are k02, k11, k12, and k22. All seven active weights sum to one.

The cyclic 2048/3072/4096 output-length schedule, curriculum.expected_budget_fraction = 0.65, and all other top8 behavior are preserved. k22 remains nonnegative and integral through observation processing.

The AR(1), AR(2), and threshold-AR recurrences use deterministic Numba kernels from proven_speed_up.md. This is a CPU-generation optimization only: fastmath and parallel execution are disabled, RNG remains outside the compiled kernels, and fixed-seed generated corpus bytes are unchanged.

Files

7 items
  • generator.py

    63e638ad0987

    39.6 KB

  • tests/test_variable_lengh.py

    36382ff2085e

    6.5 KB

  • tests/test_exact_recurrence_kernels.py

    b3918bf48cbf

    3.8 KB

  • README.md

    7b926b73f7ce

    869 B

  • config.json

    bbc5a406e1e6

    801 B

  • VARIABLE_LENGTH_UPDATE.md

    c12601e50f37

    460 B

  • requirements.txt

    101d40006778

    279 B