Downloads · 30 days
12
12% of all-time downloads
ryankim17920/qwen3p5-2b-luspo-convex-muon-lr1e4
qwen3p5-2b-luspo-convex-muon-lr1e4 is a text generation model from ryankim17920. Use it when you need the model to write or continue text. The card lists the license as apache-2.0.
vs base Qwen3.5-2B: InD acc 83.8→71.4, total output tokens 3240→383 (-88%)
Downloads · 30 days
12
12% of all-time downloads
All-time downloads
97
Public
Parameters
2.2B
16.8 GB on disk
Likes
0
Public
Click a slice to open those files.
.pt12.3 GB · 58%
From the Hugging Face model README
vs base Qwen3.5-2B: InD acc 83.8→71.4, total output tokens 3240→383 (-88%)
gpqa_diamond (OOD) acc 7.6→27.1 (+19.5 pp, +258%)
Trained via GRPO with luspo loss, convex_mean reward shape (alpha=0.1), muon optimizer, lr=2.0e-06, G=8, max_steps=200, max_completion_length=8000, evaluated over 3 seeds.
| Dataset | Base | Tuned (mean ± std) | Δ (pp, rel %) |
|---|---|---|---|
| gsm8k | 81.2 | 38.5 ± 10.0 | -42.7 pp, -53% |
| arc_challenge | 85.2 | 76.3 ± 0.8 | -8.8 pp, -10% |
| arc_easy | 97.7 | 91.7 ± 1.0 | -6.0 pp, -6% |
| commonsenseqa | 69.8 | 66.5 ± 0.9 | -3.3 pp, -5% |
| openbookqa | 82.3 | 68.5 ± 2.6 | -13.8 pp, -17% |
| qasc | 76.5 | 67.5 ± 0.0 | -9.0 pp, -12% |
| sciq | 93.8 | 91.0 ± 1.0 | -2.8 pp, -3% |
| mmlu_pro(OOD) | 33.8 | 24.5 ± 1.8 | -9.3 pp, -28% |
| mmlu_redux(OOD) | 52.5 | 44.5 ± 1.3 | -8.0 pp, -15% |
| gpqa_diamond(OOD) | 7.6 | 27.1 ± 0.6 | +19.5 pp, +258% |
| InD Average | 83.8 | 71.4 ± 1.9 | -12.4 pp, -15% |
| OOD | 31.4 | 32.1 ± 0.9 | +0.7 pp, +2% |
| ALL | 68.1 | 59.6 ± 1.5 | -8.5 pp, -12% |
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).
| Dataset | Base | Tuned (mean ± std) | Reduction % |
|---|---|---|---|
| gsm8k | 4450 | 2145 ± 478 | -52% |
| arc_challenge | 3157 | 106 ± 60 | -97% |
| arc_easy | 1871 | 78 ± 23 | -96% |
| commonsenseqa | 3949 | 161 ± 50 | -96% |
| openbookqa | 3378 | 148 ± 102 | -96% |
| qasc | 3932 | 24 ± 2 | -99% |
| sciq | 1944 | 22 ± 0 | -99% |
| mmlu_pro(OOD) | 6582 | 289 ± 57 | -96% |
| mmlu_redux(OOD) | 5589 | 192 ± 59 | -97% |
| gpqa_diamond(OOD) | 8001 | 395 ± 23 | -95% |
| InD Average | 3240 | 383 ± 67 | -88% |
| OOD | 6720 | 292 ± 41 | -96% |
| ALL | 4281 | 356 ± 35 | -92% |
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).