Weights-only best-of snapshot from a "fresh-optim restart" experiment: resumed from mixed3-v3-beta004 best1-step5500 with fresh Adam state (no optimizer resume). Wiping optimizer momentum unlocked ~+5pp IF on top of the already-strong base.

Same base config (β=0.004, 3-way mix, from v31 SFT), just fresh optim.

Full offline eval

eval	v31 SFT	v3 step-5500	freshopt-step250
IFEval-DA p-strict	21.2	32.7	38.0
IFEval-DA p-loose	22.0	34.0	39.1
IFEval-DA i-strict	35.2	48.8	54.0
IFEval-DA i-loose	35.8	49.9	54.9
GSM8K pass@1 (n=1317)	17.4	27.9	28.5
SciQ open-Q (n=1000)	13.5	14.1	14.3
CIT-gen (n=720)	29.9	29.2	29.2
textman_summary chrF++	41.1	40.9	41.1
textman_rewrite chrF++	46.5	48.2	48.0
CITMC (n=720)	48.2	47.8	48.1
SciQ-MC (n=1000)	—	58.7	58.5
PIQA (n=100)	53	57.0	57.0
ARC-Easy chat-MC	44.4	42.1	42.2
ARC-Challenge chat-MC	29.4	29.1	29.3
ARC-Easy logp	40.6	41.4	41.2
ARC-Challenge logp	27.5	26.8	26.7
GPQA-Diamond chat-MC	—	21.2	21.2
GPQA-Diamond logp	—	23.7	24.2
IFBench-DA p-strict	—	8.7	8.0
IFBench-DA i-strict	—	9.0	8.7

Composite (from in-training callback, all full-test): 2.183 (step-5500 was 2.101).

Load

python

from transformers import AutoModelForCausalLM, AutoTokenizer
REPO = "jensjepsen/danish-lm-400m-grpo-mixed3-v3-beta004-freshopt-best-step250"
model = AutoModelForCausalLM.from_pretrained(REPO)
tok = AutoTokenizer.from_pretrained(REPO)

Chat: <|user|>{q}<|end|><|assistant|>, stop on <|end|>.

Weights only.