OPST Core Ablation OPST - MMLU: leaderboard
OPST v1: Gemma 4 E2B/E4B Base means instruction-tuned; all training methods are separate checkpoints. OPST evaluates EMA teacher, not student. Nonthinking multiple-choice uses answer-option log-likelihood (MMLU 5-shot, other MC zero-shot, length normalization per lm-eval default); thinking parses final Answer: X. GSM8K 8-shot CoT exact numerical match; MATH-500 symbolic equivalence. Training topics and supervision preserved; no unpublished checkpoint URL or decoding settings inferred. Wrong-answer robustness experiment; both filters disabled; separate poisoning fractions. MMLU. Metric: Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | OPST Gemma 4 E2B OPST wrong-answer-fraction=0% | 42.7 | |
| 2 | OPST Gemma 4 E2B OPST wrong-answer-fraction=25% | 42.1 | |
| 3 | OPST Gemma 4 E2B OPST wrong-answer-fraction=50% | 41.4 | |
| 4 | OPST Gemma 4 E2B OPST wrong-answer-fraction=75% | 40.8 |
Interactive version: theaggregate.ai/benchmark?slug=opst-core-ablation-opst-mmlu · How It Works · Data refreshed daily, snapshot 2026-10-27.