LoRA adapter (r = 16, alpha = 32) for Qwen/Qwen3-4B-Base, trained with GRPO on the Countdown numbers game with a Python tool in
the loop: recipe v2, demo prompt, seed 2, +45 steps. The policy writes <think>, may run <python> (its stdout comes back in <output>), and
answers in <answer>. Training code, configs, logs and the write-up: the accompanying repository
(configs/experiments/qwen3-4b-v2-demo-s2-cont/, REPORT.md, EXPERIMENTS.md).
holdout3 (1,000 problems generated after all model decisions, greedy, 768 tokens, two code rounds): 68.2% pass@1 (easy 87.8%, medium 57.0%, hard 53.3%).
Trained with the fewshot_multi_retry prompt; evaluate with the same prompt, since the policies are prompt-specific.
Bucket run ids: q4v2s2long (step 45).
hf download prada5023/countdown-rl-qwen3-4b-v2-demo-s2-cont --local-dir adapters/qwen3-4b-v2-demo-s2-cont
uv run countdown-eval --model Qwen/Qwen3-4B-Base --adapter adapters/qwen3-4b-v2-demo-s2-cont --eval-file data/holdout3.jsonl \
--prompt-style fewshot_multi_retry --max-new-tokens 768 --max-rounds 2 --attn sdpa_gqa
Qwen/Qwen3-4B-Base