CompressedMichael/Qwen2.5-3B-Instruct-SparseGPT-50pct

Qwen2.5-3B-Instruct — SparseGPT 50%

Parent: Qwen/Qwen2.5-3B-Instruct, revision aa8e72537993ba99e69dfaafa59ed015b17504d1.

50% unstructured sparsity in transformer linear weights; BF16 checkpoint. Calibration: 128 C4 training windows, 2,048 tokens, seed 0. Embeddings and normalization weights are unchanged from Instruct. See ../provenance.json and ../validation.json.

GSM8K (4-shot chat, greedy, 4,096-token cap): 47.46% (626/1319). See results/gsm8kqwen25vllm4shotchatqwen253binssparsegpt50instructparent20260909fs4g4096ctx8192seed0/_homechienPruningGBLM-Pruneroutqwen253binstruct5020260909_sparsegptmodel/results2026-09-09T15-46-03.788488.json in /home/chien/Pruning.