ryankim17920/qwen3p5-2b-luspo-rank_a10-adamw-lr1e6

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

vs base Qwen3.5-2B: InD acc 83.8→75.4, total output tokens 3240→101 (-97%)

gpqa_diamond (OOD) acc 7.6→33.5 (+25.9 pp, +342%)

Trained via GRPO with luspo loss, rank reward shape (alpha=0.1), 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	35.0 ± 0.9	-46.2 pp, -57%
arc_challenge	85.2	82.5 ± 1.5	-2.7 pp, -3%
arc_easy	97.7	93.5 ± 0.0	-4.2 pp, -4%
commonsenseqa	69.8	71.0 ± 1.0	+1.2 pp, +2%
openbookqa	82.3	76.7 ± 1.6	-5.7 pp, -7%
qasc	76.5	76.2 ± 0.8	-0.3 pp, -0%
sciq	93.8	93.0 ± 0.0	-0.8 pp, -1%
mmlu_pro(OOD)	33.8	26.2 ± 2.0	-7.7 pp, -23%
mmlu_redux(OOD)	52.5	49.8 ± 2.8	-2.7 pp, -5%
gpqa_diamond(OOD)	7.6	33.5 ± 1.5	+25.9 pp, +342%
InD Average	83.8	75.4 ± 0.0	-8.4 pp, -10%
OOD	31.4	36.5 ± 1.6	+5.1 pp, +16%
ALL	68.1	63.8 ± 0.5	-4.3 pp, -6%

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	156 ± 4	-96%
arc_challenge	3157	94 ± 1	-97%
arc_easy	1871	94 ± 0	-95%
commonsenseqa	3949	85 ± 2	-98%
openbookqa	3378	88 ± 1	-97%
qasc	3932	100 ± 2	-97%
sciq	1944	93 ± 2	-95%
mmlu_pro(OOD)	6582	100 ± 3	-98%
mmlu_redux(OOD)	5589	101 ± 2	-98%
gpqa_diamond(OOD)	8001	108 ± 0	-99%
InD Average	3240	101 ± 1	-97%
OOD	6720	103 ± 1	-98%
ALL	4281	102 ± 0	-98%

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).