Qwen3.5-2B - luspo/rank (adamw)

vs base Qwen3.5-2B: InD acc 83.8→78.3, total output tokens 3240→223 (-93%)

gpqa_diamond (OOD) acc 7.6→31.6 (+24.1 pp, +318%)

Trained via GRPO with luspo loss, rank reward shape (alpha=0.15), 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 54.3 ± 4.0 -26.8 pp, -33%
arc_challenge 85.2 79.2 ± 0.8 -6.0 pp, -7%
arc_easy 97.7 94.2 ± 1.0 -3.5 pp, -4%
commonsenseqa 69.8 73.7 ± 3.3 +3.8 pp, +5%
openbookqa 82.3 77.7 ± 0.8 -4.7 pp, -6%
qasc 76.5 76.0 ± 0.5 -0.5 pp, -1%
sciq 93.8 92.8 ± 1.2 -1.0 pp, -1%
mmlu_pro(OOD) 33.8 31.8 ± 1.9 -2.0 pp, -6%
mmlu_redux(OOD) 52.5 54.0 ± 4.4 +1.5 pp, +3%
gpqa_diamond(OOD) 7.6 31.6 ± 3.6 +24.1 pp, +318%
InD Average 83.8 78.3 ± 0.3 -5.5 pp, -7%
OOD 31.4 39.2 ± 0.7 +7.8 pp, +25%
ALL 68.1 66.6 ± 0.1 -1.5 pp, -2%

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 410 ± 6 -91%
arc_challenge 3157 196 ± 1 -94%
arc_easy 1871 192 ± 2 -90%
commonsenseqa 3949 183 ± 1 -95%
openbookqa 3378 183 ± 1 -95%
qasc 3932 207 ± 1 -95%
sciq 1944 194 ± 3 -90%
mmlu_pro(OOD) 6582 277 ± 2 -96%
mmlu_redux(OOD) 5589 269 ± 25 -95%
gpqa_diamond(OOD) 8001 289 ± 1 -96%
InD Average 3240 223 ± 1 -93%
OOD 6720 278 ± 9 -96%
ALL 4281 240 ± 3 -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
Safetensors
Model size
2B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ryankim17920/qwen3p5-2b-luspo-rank_a15-adamw-lr1e6

Finetuned
Qwen/Qwen3.5-2B
Finetuned
(325)
this model