Datasets:
metric = weighted composite (accuracy 0.5 + reasoning-cipher-adherence 0.35 + brevity 0.15), tracked alongside raw accuracy and raw adherence separately direction = higher compute_cap = single shared A6000, target ~4-6 GPU-hours for this round (shared box, be a good citizen)
--- GRPO run config ---
base_model = Qwen/Qwen3-4B-Instruct-2507 method = LoRA (r=16, alpha=32) + GRPOTrainer (TRL 1.7.1), no vLLM (HF generate backend) num_generations = 8 # rollouts per prompt group per_device_batch = 4 grad_accum = 4 max_completion_length = 512 lr = 2e-6 beta (KL) = 0.04 # nonzero per reward-hacking literature (DEEPRESEARCH.md) curriculum = Phase 0: correctness+format reward only, ~40 steps warmup Phase 1: add cipher-adherence + brevity reward, remaining steps max_steps = 300 (Phase 0: 40, Phase 1: 260) -- reassess after Phase 1's first ~100 steps show a clear adherence trend or plateau --- spent --- steps_run = 0 gpu_hours_used = 0 best_adherence = 0 (baseline: 0, model never trained on this before) best_accuracy = n/a (to be measured pre-training as baseline on the same held-out set) champion = none yet