Downloads · 30 days
12
12% of all-time downloads
ryankim17920/qwen3p5-2b-luspo-convex-adamw-lr1e6
qwen3p5-2b-luspo-convex-adamw-lr1e6 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→76.9, total output tokens 3240→315 (-90%)
Downloads · 30 days
12
12% of all-time downloads
All-time downloads
100
Public
Parameters
2.2B
23.9 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors17.7 GB · 54%
From the Hugging Face model README
vs base Qwen3.5-2B: InD acc 83.8→76.9, total output tokens 3240→315 (-90%)
gpqa_diamond (OOD) acc 7.6→28.6 (+21.0 pp, +278%)
Trained via GRPO with luspo loss, convex_mean reward shape (alpha=0.1), adamw optimizer, lr=1.0e-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 | 61.2 ± 20.8 | -20.0 pp, -25% |
| arc_challenge | 85.2 | 78.2 ± 1.3 | -7.0 pp, -8% |
| arc_easy | 97.7 | 91.7 ± 0.8 | -6.0 pp, -6% |
| commonsenseqa | 69.8 | 68.8 ± 1.9 | -1.0 pp, -1% |
| openbookqa | 82.3 | 74.8 ± 1.6 | -7.5 pp, -9% |
| qasc | 76.5 | 72.5 ± 0.9 | -4.0 pp, -5% |
| sciq | 93.8 | 91.3 ± 0.6 | -2.5 pp, -3% |
| mmlu_pro(OOD) | 33.8 | 30.3 ± 4.3 | -3.5 pp, -10% |
| mmlu_redux(OOD) | 52.5 | 53.3 ± 3.3 | +0.8 pp, +2% |
| gpqa_diamond(OOD) | 7.6 | 28.6 ± 2.4 | +21.0 pp, +278% |
| InD Average | 83.8 | 76.9 ± 2.9 | -6.9 pp, -8% |
| OOD | 31.4 | 37.5 ± 0.6 | +6.1 pp, +19% |
| ALL | 68.1 | 65.1 ± 1.9 | -3.0 pp, -4% |
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 | 1445 ± 161 | -68% |
| arc_challenge | 3157 | 134 ± 5 | -96% |
| arc_easy | 1871 | 124 ± 0 | -93% |
| commonsenseqa | 3949 | 122 ± 4 | -97% |
| openbookqa | 3378 | 130 ± 12 | -96% |
| qasc | 3932 | 136 ± 11 | -97% |
| sciq | 1944 | 116 ± 3 | -94% |
| mmlu_pro(OOD) | 6582 | 425 ± 393 | -94% |
| mmlu_redux(OOD) | 5589 | 310 ± 214 | -94% |
| gpqa_diamond(OOD) | 8001 | 365 ± 282 | -95% |
| InD Average | 3240 | 315 ± 23 | -90% |
| OOD | 6720 | 367 ± 296 | -95% |
| ALL | 4281 | 331 ± 92 | -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).