Qwen3.5-2B - luspo/min_anchor (adamw)
vs base Qwen3.5-2B: InD acc 83.8→79.0, total output tokens 3240→203 (-94%)
gpqa_diamond (OOD) acc 7.6→28.3 (+20.7 pp, +273%)
Trained via GRPO with luspo loss, min_anchor reward shape (alpha=0.05), adamw optimizer, lr=1.0e-06, G=8, max_steps=200, max_completion_length=8000, evaluated over 3 seeds.
Accuracy vs base Qwen3.5-2B
| Dataset | Base | Tuned (mean ± std) | Δ (pp, rel %) |
|---|---|---|---|
| gsm8k | 81.2 | 57.7 ± 6.8 | -23.5 pp, -29% |
| arc_challenge | 85.2 | 81.3 ± 1.3 | -3.8 pp, -5% |
| arc_easy | 97.7 | 95.3 ± 1.2 | -2.3 pp, -2% |
| commonsenseqa | 69.8 | 72.3 ± 1.0 | +2.5 pp, +4% |
| openbookqa | 82.3 | 78.5 ± 3.0 | -3.8 pp, -5% |
| qasc | 76.5 | 75.7 ± 0.6 | -0.8 pp, -1% |
| sciq | 93.8 | 92.2 ± 1.6 | -1.7 pp, -2% |
| mmlu_pro(OOD) | 33.8 | 36.5 ± 0.5 | +2.7 pp, +8% |
| mmlu_redux(OOD) | 52.5 | 52.7 ± 1.3 | +0.2 pp, +0% |
| gpqa_diamond(OOD) | 7.6 | 28.3 ± 1.5 | +20.7 pp, +273% |
| InD Average | 83.8 | 79.0 ± 1.0 | -4.8 pp, -6% |
| OOD | 31.4 | 39.2 ± 0.9 | +7.8 pp, +25% |
| ALL | 68.1 | 67.1 ± 0.8 | -1.0 pp, -1% |
Δ shows the absolute change in accuracy points (pp) and the relative percent change (tuned − base) / base × 100 (rel %, shown as n/a when base accuracy is 0).
Output tokens (total) vs base Qwen3.5-2B
| Dataset | Base | Tuned (mean ± std) | Reduction % |
|---|---|---|---|
| gsm8k | 4450 | 381 ± 37 | -91% |
| arc_challenge | 3157 | 178 ± 6 | -94% |
| arc_easy | 1871 | 172 ± 5 | -91% |
| commonsenseqa | 3949 | 169 ± 2 | -96% |
| openbookqa | 3378 | 159 ± 3 | -95% |
| qasc | 3932 | 201 ± 5 | -95% |
| sciq | 1944 | 162 ± 4 | -92% |
| mmlu_pro(OOD) | 6582 | 328 ± 28 | -95% |
| mmlu_redux(OOD) | 5589 | 274 ± 7 | -95% |
| gpqa_diamond(OOD) | 8001 | 342 ± 49 | -96% |
| InD Average | 3240 | 203 ± 6 | -94% |
| OOD | 6720 | 315 ± 23 | -95% |
| ALL | 4281 | 237 ± 11 | -94% |
Output tokens = total generated tokens (full completion), 3-seed mean.
Reduction = percentage decrease in mean output tokens vs base Qwen3.5-2B (negative reduction, i.e. +, means the tuned model generates more tokens).
- Downloads last month
- 4