rl: pin TRITON_ATTN backend (FLASHINFER silently breaks vllm-lens injection on Blackwell) 28ad77d ceselder Claude Fable 5 commited on 3 days ago
rl: true per-token entropy bonus (maximize r + beta*H, explicit diversity knob, no KL) 73adae4 ceselder Claude Fable 5 commited on 3 days ago
maxact-fast: hardening fixes (6 bugs) + Dr.GRPO vllm-lens rl.py + sbatch; Probe = universal direction 48db85f ceselder Claude Fable 5 commited on 3 days ago