A best-of-N specialist variant of hemlang/Hemlock-Apothecary-7B: the epoch-3 checkpoint of an execution-reward GRPO run (grimoire / hemlock-rl) on hemlang/hemlock-codex3-SFT.

Know the trade-off (hembench zero-shot, n=5):

Mid-training RL dispersed the policy: lower first-try reliability, but the highest any-shot solve rate of any Hemlock model measured to date. Use the base Apothecary for single-shot generation; use this model when sampling several candidates and verifying with the interpreter — e.g. hembot --retry — where solve-rate within k samples is what matters. Notable per-level: systems/concurrency L4 pass@1 4/7 → 5/7, debugging L6 pass@5 3/5 → 4/5.

GRPO with LoRA r=16, β=0.1 KL, G=8, dynamic sampling, 267 deduplicated verified prompts, constant lr, checkpoint at epoch 3 of 4 (the epoch-4 endpoint re-concentrates and loses the pass@5 advantage). Q8_0 GGUF included.