countdown-rl-gemma4-e2b-v2-zeroshot-s0
LoRA adapter (r = 16, alpha = 32) for google/gemma-4-E2B (text decoder), trained with GRPO on the Countdown numbers game with a Python tool in
the loop: recipe v2, zero-shot prompt, seed 0. 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/gemma4-e2b-v2-zeroshot-s0/, REPORT.md, EXPERIMENTS.md).
holdout3 (1,000 problems generated after all model decisions, greedy, 768 tokens, two code rounds): 4.4% pass@1 (easy 9.0%, medium 1.7%, hard 1.0%).
Trained with the r1zero_tool prompt; evaluate with the same prompt, since the policies are prompt-specific.
Bucket run ids: e2v2z0-a..d.
The adapter targets the text decoder of Gemma-4-E2B, extracted from the multimodal checkpoint. Create it first (logits match the full model exactly with eager attention), and use --attn eager:
uv run python scripts/prepare_gemma4_text.py --model google/gemma-4-E2B --out models/gemma-4-E2B-text
Usage
hf download prada5023/countdown-rl-gemma4-e2b-v2-zeroshot-s0 --local-dir adapters/gemma4-e2b-v2-zeroshot-s0
uv run countdown-eval --model models/gemma-4-E2B-text --adapter adapters/gemma4-e2b-v2-zeroshot-s0 --eval-file data/holdout3.jsonl \
--prompt-style r1zero_tool --max-new-tokens 768 --max-rounds 2 --attn eager