UCL-CSSB/PlasmidRL-ICML/rejection_sampling_v2/best_of_16 - Storage Bucket

best_of_16_v2

Best-of-16 selection at per-model optimal temperature, replacing the paper-tab:baselines best-of-16 numbers (which were at the suboptimal T=0.95 for all three models). Same protocol as v1 baselines/best_of_16/: 16K candidates per cell (8K per prompt × 2 prompts), keep argmax(reward) per 16-tuple, score the 1000 selected with the analysis2 strict QC pipeline.

Generated: 2026-05-01T10:12:26.691423Z

Per-cell results (analysis2 strict QC)

Cell	Model	T	n_selected	pass rate	sha256(outputs.csv)
Base_t1	UCL-CSSB/PlasmidGPT	1.0	1000	34.9%	`93e620f4dd7179ab…`
GRPO_t1.15	UCL-CSSB/PlasmidGPT-GRPO	1.15	1000	99.6%	`c82587be8b5c564e…`
SFT_t1	UCL-CSSB/PlasmidGPT-SFT	1.0	1000	32.4%	`5a8e9f1502b786cb…`

v1 cross-check

All v2 outputs.csv SHAs are distinct from v1 baselines/best_of_16/{Base,SFT,GRPO}:

• Base: `9dc5e7aab90a15ca…`
• SFT: `c96bdf4af04ce868…`
• GRPO: `99527f852fcbe3f1…`