ryankim17920/qwen3p5-2b-luspo-rank_a00-adamw-lr1e6

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

vs base Qwen3.5-2B: InD acc 83.8→81.7, total output tokens 3240→1133 (-65%)

gpqa_diamond (OOD) acc 7.6→37.4 (+29.8 pp, +393%)

Trained via GRPO with luspo loss, rank reward shape (alpha=0.0), 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	66.7 ± 0.3	-14.5 pp, -18%
arc_challenge	85.2	83.0 ± 2.2	-2.2 pp, -3%
arc_easy	97.7	96.5 ± 0.5	-1.2 pp, -1%
commonsenseqa	69.8	72.8 ± 0.6	+3.0 pp, +4%
openbookqa	82.3	81.8 ± 1.0	-0.5 pp, -1%
qasc	76.5	78.5 ± 0.9	+2.0 pp, +3%
sciq	93.8	92.8 ± 1.3	-1.0 pp, -1%
mmlu_pro(OOD)	33.8	41.8 ± 4.0	+8.0 pp, +24%
mmlu_redux(OOD)	52.5	60.0 ± 3.0	+7.5 pp, +14%
gpqa_diamond(OOD)	7.6	37.4 ± 3.5	+29.8 pp, +393%
InD Average	83.8	81.7 ± 0.4	-2.0 pp, -2%
OOD	31.4	46.4 ± 3.1	+15.1 pp, +48%
ALL	68.1	71.2 ± 1.1	+3.1 pp, +5%

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	2347 ± 173	-47%
arc_challenge	3157	1002 ± 35	-68%
arc_easy	1871	892 ± 50	-52%
commonsenseqa	3949	955 ± 128	-76%
openbookqa	3378	881 ± 48	-74%
qasc	3932	1029 ± 10	-74%
sciq	1944	824 ± 57	-58%
mmlu_pro(OOD)	6582	1539 ± 42	-77%
mmlu_redux(OOD)	5589	1300 ± 34	-77%
gpqa_diamond(OOD)	8001	1537 ± 77	-81%
InD Average	3240	1133 ± 35	-65%
OOD	6720	1458 ± 12	-78%
ALL	4281	1230 ± 27	-71%

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).