TD-MoE Tucker decomposition on Qwen3-30B-A3B, compression matched to MoBE/Sparse-MoBE b32 (37.5% removed per gate/up matrix, 62.5% kept). Expert-preserving.

Method: td_moe
Base model: Qwen/Qwen3-30B-A3B
Compressed: gate_proj, up_proj over layers 0-48 (down_proj kept dense)
Settings: TD-MoE Tucker, compression_ratio=0.375 (removed), expert_mode=preserve

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	0.3311
arc_challenge	acc_stderr,none	0.0138
arc_challenge	acc_norm,none	0.3626
arc_challenge	acc_norm_stderr,none	0.0140
arc_easy	sample_len	2376.0000
arc_easy	acc,none	0.5724
arc_easy	acc_stderr,none	0.0102
arc_easy	acc_norm,none	0.5101
arc_easy	acc_norm_stderr,none	0.0103
winogrande	sample_len	1267.0000
winogrande	acc,none	0.6259
winogrande	acc_stderr,none	0.0136
piqa	sample_len	1838.0000
piqa	acc,none	0.6779
piqa	acc_stderr,none	0.0109
piqa	acc_norm,none	0.7051
piqa	acc_norm_stderr,none	0.0106
hellaswag	sample_len	10042.0000
hellaswag	acc,none	0.3995
hellaswag	acc_stderr,none	0.0049
hellaswag	acc_norm,none	0.5269
hellaswag	acc_norm_stderr,none	0.0050
openbookqa	sample_len	500.0000
openbookqa	acc,none	0.2680
openbookqa	acc_stderr,none	0.0198
openbookqa	acc_norm,none	0.3340
openbookqa	acc_norm_stderr,none	0.0211

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "AverageMetaheuristicsEnjoyer/qwen3-30b-a3b-td-moe-ep-fix",
    trust_remote_code=True,
    torch_dtype="bfloat16",
    device_map="auto",
)
tok = AutoTokenizer.from_pretrained("AverageMetaheuristicsEnjoyer/qwen3-30b-a3b-td-moe-ep-fix", trust_remote_code=True)

The factored experts are reconstructed at runtime via the bundled Qwen3TDMoEForCausalLM.