Buckets:

476 GB
343 files
Updated 9 days ago
README.md

LearnPruner LPM for Qwen3.8-27B. The trained LPM is BUDGET-INDEPENDENT: the training mask is argmax over the LPM's own scores, so the retention budget never enters training and is applied at inference. Use this single checkpoint for every ratio (0.0556 / 0.1111 / 0.2222 / 0.3889). Trained on LLaVA-Tuning/llava_v1_5_mix665k_qwenbox.json (mix665k with every box rewritten to Qwen's 0-1000 of the unpadded image), ALL image rows (subset_frac 1.0, not the paper's 10%), seed 2026, 1 epoch, 624,609 examples, 78,076 steps. This backbone: 4 H100s x batch 1 x 2 accumulation (effective batch 8); 5120-dim LPM on the post-merger visual tokens. Stage 2 is scored at layer 23 and applies from layer 24 of 64 (the paper's layer 12 of 32, scaled).

Total size
476 GB
Files
343
Last updated
Nov 11
Pre-warmed CDN
US EU US EU

Contributors