Language Modeling on Multi-domain Language Evaluation Suite (various)
50.5Overall ScoreRe-training (mid+post-train)†
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Re-training (mid+post-train)†Scope=mid+post-train2026.04 | 50.5 | 27 | 31.8 | 43.2 | 55.9 | 59.6 | 45.9 | 90.4 | |
| BAR 5×7B (all experts)Architecture=5×7B MoE, Expert Config=all experts2026.04 | 49.1 | 28.4 | 30.8 | 38.7 | 56.2 | 49.9 | 45.6 | 94 | |
| Re-training (post-train only)Scope=post-train only2026.04 | 47.8 | 28.6 | 31.3 | 43.9 | 48.7 | 43.6 | 45.3 | 92.8 | |
| BTX 5×7BArchitecture=5×7B MoE2026.04 | 46.7 | 23.9 | 30.6 | 36.4 | 62.1 | 32.1 | 47.9 | 93.9 | |
| Continual post-training2026.04 | 45.3 | 26.8 | 29.4 | 38.8 | 45 | 42.9 | 40.9 | 93.1 | |
| Math expert + RLArchitecture=2×7B MoE, Expert Domain=Math, Stage=RL2026.04 | 39.3 | 29 | 30.8 | 42.5 | 55.8 | 22.1 | 19.8 | 75.4 | |
| Code expert + RLArchitecture=2×7B MoE, Expert Domain=Code, Stage=RL2026.04 | 38.8 | 28.5 | 29.2 | 41 | 26.9 | 50.4 | 19.8 | 75.3 | |
| Code expert + SFTArchitecture=2×7B MoE, Expert Domain=Code, Stage=SFT2026.04 | 38.5 | 28.8 | 29.1 | 40.1 | 25.5 | 49.3 | 19.7 | 77.3 | |
| Tool use expert + SFTArchitecture=2×7B MoE, Expert Domain=Tool use, Stage=SFT2026.04 | 37.2 | 28.5 | 28.7 | 39.3 | 21.8 | 16.9 | 46.4 | 79.1 | |
| Model merging (w/o mid-train)Mid-train=Excluded2026.04 | 36.9 | 29.2 | 30 | 42.6 | 32.9 | 25.1 | 19.7 | 78.5 | |
| Math expert + SFTArchitecture=2×7B MoE, Expert Domain=Math, Stage=SFT2026.04 | 36.8 | 28.8 | 31.2 | 40.9 | 41.9 | 20.5 | 21.6 | 72.7 | |
| Safety expert + SFTArchitecture=2×7B MoE, Expert Domain=Safety, Stage=SFT2026.04 | 35.6 | 28.7 | 28.8 | 38.1 | 22.4 | 15.7 | 21.1 | 94.6 | |
| Initial model MScale=7B2026.04 | 31.3 | 28.5 | 29.8 | 48.9 | 23.6 | 11.8 | 25.3 | 51.3 | |
| Model merging (w/ mid-train)Mid-train=Included2026.04 | 6.5 | 0.1 | 10.2 | 5.9 | 0.3 | 0.4 | 19.7 | 9.1 |