Zarya-4B Release - H100 BF16 - ifeval - inst_level_strict_acc - none: leaderboard

Zarya-4B Release - H100 BF16 - ifeval - inst_level_strict_acc - none — Zarya-4B official native lm-eval results;H100, `dtype=bfloat16`, `attn_implementation=sdpa`, `apply_chat_template`, `slotted_generation=true,slot_size=16,serial_num_blocks=4,slot_threshold=0.9,token_threshold=0.4`. lm-eval0.4.13; native task version,n-shot,filter,metric,stderr retained. Hardware/dtype are separate native conditions,not separate trained models. Slotted speculative decoding,not plain masked-diffusion sampling. Template and exact decoding configuration remain explicit; source reports hardware/configuration sensitivity. Singleton observations remain native; no invented peers, scores or fit. Metric: inst_level_strict_acc (fraction).

Metric: inst_level_strict_acc (fraction). Source: huggingface.co. 1 model tracked. Updated 15 Nov 2026.

Top models

#ModelScoreOverall rank
1Zarya-4B0.6379#4590

Interactive version: theaggregate.ai/benchmark?slug=zarya-4b-release-h100-bf16-ifeval-inst-level-strict-acc-none · How It Works · Data refreshed daily, snapshot 2026-11-16.