fix(reward): use normalized ToolBench ROUGE thresholds

[MrCapricornLiu]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

The ToolBench reward returns zero even for identical plain-text action inputs. rouge supplies an F1 score in [0, 1], but ReactORM compares it against 10 and 20, so this branch always assigns zero.

Use 0.1 and 0.2 for the existing thresholds. The binary reward policy, JSON argument comparison, and tool-name checks remain unchanged.

For example, two identical Action: search\nAction Input: New York weather completions now receive 1 instead of 0; unrelated text and a different tool still receive 0.

Experiment results

• Three unittest methods cover real ROUGE scoring, string and message-dict inputs, threshold boundaries, JSON arguments, mismatched tools, and invalid inputs. The original code has four failing assertions; all tests pass with the fix.
• The real GRPO score_completions path produces [1, 1, 0, 0] for exact/overlapping/unrelated/wrong-tool completions, versus [0, 0, 0, 0] before the fix (rouge 1.0.1).
• All pre-commit hooks pass. This validates reward computation, not a full training run.

python -m unittest discover -s tests -p test_toolbench_reward.py -v