Run dir: [v2_runs]/selfhost_qwen36_05_CD45_20260424_113757_7323363
| target | program/profile | gen | n_successful | rank_key | best_i_pae | best_plddt | description |
|---|---|---|---|---|---|---|---|
| 05_CD45 | 0d487097c599 | 1 | 0 | 0.24697357 | 0.9661878 | Switch from greedy beam-search to diverse-beam-search with beam_width=8 and diversity_weight=2.0 to explore diverse binding modes that may achieve iPAE below the 0.226 threshold. | |
| 05_CD45 | 189da5b69268 | 2 | 0 | 0.24697357 | 0.9661878 | Reduce diverse-beam diversity_weight from 2.0 to 0.8 to avoid pushing search into clearly wrong interface conformations, while strengthening i_pae penalty from -1.0 to -2.0 and ... | |
| 05_CD45 | 1c536628cd91 | 3 | 4 | 0.19810465 | 0.9684078 | Enable full sequence_hallucination refinement (BindCraft-style AF2 co-design) with increased hard/greedy iterations and stronger interface pAE loss to optimize sequence-structur... | |
| 05_CD45 | 2b558efb6ac3 | 2 | 0 | 0.28966004 | 0.9482349 | Two-stage program: cheap first stage with single-pass sampling (nsteps=200, nsamples=16) for broad exploration, followed by sequence_hallucination refinement stage on top candid... | |
| 05_CD45 | 3ef9313621da | 6 | 0 | 0.28402632 | 0.9581066 | Two-stage program: cheap first stage (single-pass, 200 steps, 32 samples) for broad exploration, followed by conditional second stage (diverse-beam-search, 400 steps, sequence_h... | |
| 05_CD45 | 60fa7d6426ef | 0 | 0 | 0.24697357 | 0.9661878 | seed (balanced baseline) | |
| 05_CD45 | 75a81e6da237 | 7 | 86 | 0.14582375 | 0.9640748 | Increase diverse-beam-search beam_width from 8 to 12 and n_branch from 4 to 6 while slightly deepening i_pae penalty to -2.5, testing whether expanded beam coverage with stronge... | |
| 05_CD45 | a0c828c67564 | 6 | 29 | 0.1732057 | 0.9645294 | Switch from MCTS to diverse-beam-search to leverage the 13-success beam-search performance pattern, with moderate diversity_weight=0.8 that was tuned earlier for CD45. | |
| 05_CD45 | b5ad2bab0ea1 | 5 | 5 | 0.13042143 | 0.96784824 | Switch search to MCTS and increase simulations/samples to leverage freed compute from reduced refinement iterations, targeting improved structural diversity for CD45. | |
| 05_CD45 | bfc06ff2a57d | 4 | 13 | 0.1632463 | 0.9670205 | Reduce greedy/hard refinement iterations to 20/10 to save compute while doubling nsamples to 8 and reducing batch_size to 4 for safer memory usage, aiming to increase successful... | |
| 05_CD45 | ce1b7f45f2df | 8 | Increase nsamples from 16 to 20 to expand the per-stage sample budget while retaining the proven diverse-beam-search configuration (beam_width=12, n_branch=6, diversity_weight=0... | ||||
| 05_CD45 | d56247f6ea85 | 5 | 3 | 0.19227476 | 0.96015275 | Switch search algorithm to MCTS with balanced exploration to diversify interface | |
| 05_CD45 | e77877a97767 | 1 | 0 | Strengthen iPAE penalty to -3.0, add modest pLDDT weight 0.5 (exploratory stress-test), increase beam_width to 12, extend sampling to 600 steps, update checkpoints to match. | |||
| 05_CD45 | fb2c9edacef3 | 7 | 54 | 0.14725764 | 0.9637073 | Increase diverse-beam beam_width to 12 and stage nsamples to 20 to expand search coverage per the diversify-early hypothesis while keeping active compute within 512 limit. |
Generated 2026-04-25