Downloads · 30 days
11
12% of all-time downloads
ryankim17920/qwen3p5-2b-luspo-minanchor-adamw-lr2e6
qwen3p5-2b-luspo-minanchor-adamw-lr2e6 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→66.8, total output tokens 3240→1 (-100%)
Downloads · 30 days
11
12% of all-time downloads
All-time downloads
90
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→66.8, total output tokens 3240→1 (-100%)
gpqa_diamond (OOD) acc 7.6→31.6 (+24.1 pp, +318%)
Trained via GRPO with luspo loss, min_anchor reward shape (alpha=0.1), adamw 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 | 6.5 ± 0.5 | -74.7 pp, -92% |
| arc_challenge | 85.2 | 76.2 ± 2.6 | -9.0 pp, -11% |
| arc_easy | 97.7 | 91.7 ± 1.6 | -6.0 pp, -6% |
| commonsenseqa | 69.8 | 69.2 ± 0.8 | -0.7 pp, -1% |
| openbookqa | 82.3 | 66.7 ± 1.8 | -15.7 pp, -19% |
| qasc | 76.5 | 66.0 ± 4.4 | -10.5 pp, -14% |
| sciq | 93.8 | 91.3 ± 0.6 | -2.5 pp, -3% |
| mmlu_pro(OOD) | 33.8 | 24.5 ± 1.8 | -9.3 pp, -28% |
| mmlu_redux(OOD) | 52.5 | 47.8 ± 1.9 | -4.7 pp, -9% |
| gpqa_diamond(OOD) | 7.6 | 31.6 ± 0.8 | +24.1 pp, +318% |
| InD Average | 83.8 | 66.8 ± 1.6 | -17.0 pp, -20% |
| OOD | 31.4 | 34.7 ± 0.4 | +3.3 pp, +10% |
| ALL | 68.1 | 57.2 ± 1.2 | -10.9 pp, -16% |
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 | 1 ± 0 | -100% |
| arc_challenge | 3157 | 1 ± 0 | -100% |
| arc_easy | 1871 | 1 ± 0 | -100% |
| commonsenseqa | 3949 | 1 ± 0 | -100% |
| openbookqa | 3378 | 1 ± 0 | -100% |
| qasc | 3932 | 1 ± 0 | -100% |
| sciq | 1944 | 1 ± 0 | -100% |
| mmlu_pro(OOD) | 6582 | 1 ± 0 | -100% |
| mmlu_redux(OOD) | 5589 | 1 ± 0 | -100% |
| gpqa_diamond(OOD) | 8001 | 1 ± 0 | -100% |
| InD Average | 3240 | 1 ± 0 | -100% |
| OOD | 6720 | 1 ± 0 | -100% |
| ALL | 4281 | 1 ± 0 | -100% |
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).