MiniCPM-2B-dpo-bf16-llama-format: benchmark results
A Llama-format DPO checkpoint of MiniCPM, developed jointly by ModelBest and TsinghuaNLP. The family card quotes 2.4B parameters excluding embeddings; that figure is not the full learned parameter total. Provider: OpenBMB. Access: Open.
Aggregate Rating 1092 ± 40, rank #10799 of 10916 rated models, from 5 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard v1 - WinoGrande | 51.78 | Accuracy (%) (5-shot) | 6 |
| Open LLM Leaderboard v1 - ARC Challenge | 25.6 | Normalized accuracy (%) (25-shot) | 4.9 |
| Open LLM Leaderboard v1 - GSM8K | 0 | Accuracy (%) (5-shot) | 4.5 |
| Open LLM Leaderboard v1 - MMLU | 24.24 | Accuracy (%) (5-shot) | 3.3 |
| Open LLM Leaderboard v1 - HellaSwag | 22.42 | Normalized accuracy (%) (10-shot) | 0 |
Interactive version: theaggregate.ai/model?slug=minicpm-2b-dpo-bf16-llama-format · How It Works · Data refreshed daily, snapshot 2026-11-08.