Hallucination Mitigation on Delulu N=1,950, 7 languages (held-out)
61.8Exact Match (EM)Qwen2.5-Coder-7B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Qwen2.5-Coder-7BModel size=7B, Training Stage=SFT, Truncation strategy=first-N-lines truncation2026.06 | 61.8 | 85 | |
| Qwen2.5-Coder-7BModel size=7B, Training Stage=SFT, Truncation strategy=none2026.06 | 61.4 | 87 | |
| Qwen2.5-Coder-3BModel size=3B, Training Stage=SFT, Truncation strategy=first-N-lines truncation2026.06 | 59.2 | 83.5 | |
| Qwen2.5-Coder-3BModel size=3B, Training Stage=SFT, Truncation strategy=none2026.06 | 58.5 | 85 | |
| Qwen2.5-Coder-3BModel size=3B, Training Stage=base, Truncation strategy=first-N-lines truncation2026.06 | 52 | 79.8 | |
| Qwen2.5-Coder-7BModel size=7B, Training Stage=base, Truncation strategy=first-N-lines truncation2026.06 | 48.8 | 77 | |
| Qwen2.5-Coder-3BModel size=3B, Training Stage=base, Truncation strategy=none2026.06 | 45.7 | 72 | |
| Qwen2.5-Coder-7BModel size=7B, Training Stage=base, Truncation strategy=none2026.06 | 42.8 | 65 |