selfhost_qwen36_23_BetV1_20260424_113757_7323365

selfhost summary report

Run dir: [v2_runs]/selfhost_qwen36_23_BetV1_20260424_113757_7323365

targetprogram/profilegenn_successfulrank_keybest_i_paebest_plddtdescription
23_BetV129cf0ab15f1030Increase beam_width from 8 to 12 in stage 0 for broader early diversity, reduce nsteps to 200 with matching step_checkpoints to halve per-candidate compute in stage 0, and appen...
23_BetV12c4892e6bfa04Increase stage 0 refinement depth to n_temp=80/n_hard=12, switch refine_targets to all, and add plddt=0.5 reward weight to boost confidence and interface sharpness.
23_BetV13efdc1348845200.324391480.9182643Switch from beam-search to MCTS (n_simulations=50) to leverage the paper's finding that MCTS excels on harder targets, and increase sampling.nsteps from 400 to 600 with matching...
23_BetV14213fe45aa1a20Increase beam_width from 8 to 12 for broader diversity, increase sampling.nsteps from 400 to 600 for more thorough exploration, align step_checkpoints with new nsteps, and reduc...
23_BetV154e0a7d9fe1230Reducing stage 0 nsteps to 200 and step_checkpoints to match, increasing beam_width to 12 and nsamples to 10 for diversity in cheap exploration, and appending a conditional thir...
23_BetV1635c3ca770c7200.293039650.9283373Append a second stage conditioned on n_successful>=2; stage 0 reduces beam_width to 4 for more diversity (nsamples=16, nsteps=200), stage 1 deepens refinement (nsamples=8, nstep...
23_BetV16d9e1a413d6c30Increase beam_width to 12 and nsteps to 600 in stage 0 for broader, higher-quality exploration (with matching step_checkpoints), while raising greedy and temperature iterations ...
23_BetV16dbea9c49eab000.284821480.9414049seed (balanced baseline)
23_BetV174a1d20708754150.160202610.9444216Further strengthen stage 2 refinement by increasing n_temp_iters from 60 to 70 and n_hard_iters from 8 to 10, plus switching the stage 2 condition from n_successful >= 2 to a qu...
23_BetV1761f1bafefd3380.183904250.95171094Increase stage 0 beam_width to 16 for broader exploration (batch_size=2 keeps memory at 16×8×2=256), increase stage 1 nsteps to 600 with matching step_checkpoints and n_temp_ite...
23_BetV1783ccca607e910Add i_ptm (+0.5) and i_con (+1.0) reward weights from BindCraft's filtering criteria and increase i_pae penalty from -1.0 to -2.0 to push harder on interface accuracy, plus wide...
23_BetV191062ca806d430Increase beam_width from 8 to 12 and nsamples from 8 to 10 in stage 0 for broader early diversity, halve nsteps from 400 to 200 for cheaper per-sample exploration (12×10×4=480 ≤...
23_BetV1c20523dd958f3180.15987060.94889796Strengthen refinement by increasing n_temp_iters from 45 to 60 and n_hard_iters from 5 to 8, plus boosting i_pae loss weight from 0.1 to 0.3 to align with the strong -2.0 i_pae ...
23_BetV1caac57428874160.148742820.9489059Strengthen i_pae penalty to -2.0 and add contact/interface-contact reward terms (con=0.5, i_con=0.5) to guide interface formation, while enabling sequence_hallucination refineme...
23_BetV1fac00cffd93a2110.15987450.93689823Increase nsamples from 4 to 8 for broader exploration (doubling candidates), halve batch_size from 8 to 4 to keep memory at 8×8×4=256, and append a second refinement stage (clon...

Generated 2026-04-25