Axiom JAX/MXFP4 pilot — now with the MFU-55 research card and its trace dataset: 452 → 1,203.7 tok/s on one DGX Spark in 8 days.
Roman Nekrasov
Rob1234567
AI & ML interests
Areas of interest: agentic mid-training, reinforcement learning with reward verification (RLVR), scaling agent environments, interleaved agent reasoning with tools
Recent Activity
updated a Space 4 days ago
Rob1234567/personal-manifesto published a Space 4 days ago
Rob1234567/personal-manifesto updated a collection 9 days ago
AxiomOrganizations
None yet
Axiom JAX/MXFP4 pilot — now with the MFU-55 research card and its trace dataset: 452 → 1,203.7 tok/s on one DGX Spark in 8 days.
Хирургия весов Qwen2.5-0.5B → гибрид 3:1 KDA/GQA: CPT, recall-серии и экономика длинного контекста (память ×3.9, decode ×3.65 на 512K, ≤32K).
Qwen3.5-4B, дообучен на 124K концептов arXiv. safetensors (GRPO v10) + GGUF Q4_K_M с PDF-отчётами и кейсами.