llada-8b-base-sparsegpt-30

Source: GSAI-ML/LLaDA-8B-Base at 0f2787f2d87eac5eed8a087d5ecd24277e6255b2. Structured pruning target: 30%; pattern: 3:10. Measured values are in pruning_statistics.json. The weights and tokenizer were saved and verified by a forward pass after reload. Benchmark evaluation has not been run.

Load with transformers.AutoModel.from_pretrained(repo_id, trust_remote_code=True).

Pruned checkpoint evaluation

Full completed benchmarks only. Original weight revisions are pinned below; updating this model card does not change evaluated weights. Settings: BF16, seed 42, MC=128, batch size 32. LLaDA generation uses 256 steps and 1024 output tokens; GSM8K uses its built-in 5-shot setting, other tasks are zero-shot.

Benchmark	Metric	Value	Weights	Raw JSON
task:arc_challenge	acc,none	0.453072	d4eb47ef5bb6	JSON
task:arc_challenge	acc_norm,none	0.485495	d4eb47ef5bb6	JSON
task:arc_easy	acc,none	0.749158	d4eb47ef5bb6	JSON
task:arc_easy	acc_norm,none	0.743687	d4eb47ef5bb6	JSON
task:winogrande	acc,none	0.686661	d4eb47ef5bb6	JSON

An absent benchmark is pending, not a zero score. Individual raw files record the complete evaluation configuration and checkpoint provenance.