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 v1, phase d without hinted resampling (ablation). 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-v1-handtuned-nohint/, REPORT.md, EXPERIMENTS.md).
Ablation checkpoint; not evaluated on holdout3 (see EXPERIMENTS.md, section 2).
Trained with the fewshot_multi_retry prompt; evaluate with the same prompt, since the policies are prompt-specific.
Bucket run ids: p10bnh (step 80).
hf download prada5023/countdown-rl-qwen3-4b-v1-handtuned-nohint --local-dir adapters/qwen3-4b-v1-handtuned-nohint
uv run countdown-eval --model Qwen/Qwen3-4B-Base --adapter adapters/qwen3-4b-v1-handtuned-nohint --eval-file data/holdout3.jsonl \
--prompt-style fewshot_multi_retry --max-new-tokens 768 --max-rounds 2 --attn sdpa_gqa
Qwen/Qwen3-4B-Base