Sparse-MoBE stripped of LISTA priors (T=0, S=0, square encoder, random w-init). Head-to-head vs MoBE b32 at identical optimization (lr=0.07): isolates the W_e+theta coefficient reparameterization, which shares MoBE's function class. See docs/sparse_mobe_mechanism.md.

Method: sparse_mobe
Base model: Qwen/Qwen3-30B-A3B
Compressed: gate_proj, up_proj over layers 0-48 (down_proj kept dense)
Settings: sparse_mobe, num_B=32, activation=tanh

See config.toml and training.json for the full run configuration and per-layer reconstruction stats.

Evaluation

Task	Metric	Value
arc_challenge	sample_len	1172.0000
arc_challenge	acc,none	0.5316
arc_challenge	acc_stderr,none	0.0146
arc_challenge	acc_norm,none	0.5657
arc_challenge	acc_norm_stderr,none	0.0145
arc_easy	sample_len	2376.0000
arc_easy	acc,none	0.7963
arc_easy	acc_stderr,none	0.0083
arc_easy	acc_norm,none	0.7938
arc_easy	acc_norm_stderr,none	0.0083
winogrande	sample_len	1267.0000
winogrande	acc,none	0.7127
winogrande	acc_stderr,none	0.0127
piqa	sample_len	1838.0000
piqa	acc,none	0.7856
piqa	acc_stderr,none	0.0096
piqa	acc_norm,none	0.7954
piqa	acc_norm_stderr,none	0.0094
hellaswag	sample_len	10042.0000
hellaswag	acc,none	0.5663
hellaswag	acc_stderr,none	0.0049
hellaswag	acc_norm,none	0.7446
hellaswag	acc_norm_stderr,none	0.0044
openbookqa	sample_len	500.0000
openbookqa	acc,none	0.3240
openbookqa	acc_stderr,none	0.0210
openbookqa	acc_norm,none	0.4500
openbookqa	acc_norm_stderr,none	0.0223

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "AverageMetaheuristicsEnjoyer/qwen3-30b-a3b-sparse-mobe-b32-minimal-h100",
    trust_remote_code=True,
    torch_dtype="bfloat16",
    device_map="auto",
)
tok = AutoTokenizer.from_pretrained("AverageMetaheuristicsEnjoyer/qwen3-30b-a3b-sparse-mobe-b32-minimal-h100", trust_remote_code=True)

The factored experts are reconstructed at runtime via the bundled Qwen3SparseMoBEForCausalLM.