Is tool_call_reward expected to be a fixed -0.1 ?

[Xinm123]

Hi
It seems that extra_info in the DAPO Manager doesn't include the num_turns key, causing tool_call_reward = (num_turns - 2) / 2 * 0.1 is 0.1 when result["score"] < 0 in reward.py

[koriyoshi2041]

Traced this: the constant -0.1 comes from num_turns defaulting to 0. recipe/demystify/reward.py reads extra_info.get("num_turns", 0), and naive.py populates that key from the batch (extra_info["num_turns"] = data_item.non_tensor_batch.get("__num_turns__")), but dapo.py never did — and the demystify scripts run with reward_manager=dapo. So (num_turns - 2) / 2 * 0.1 is always (0-2)/2*0.1 = -0.1.

Opened #15 to mirror the two num_turns lines from naive.py into dapo.py.