fix(cookbooks): normalize LaTeX and markdown formatting in Fireworks …

[shobhitagnihotri69]

Summary

The arithmetic task in cookbooks/fireworks-rl-training checks the final integer on the last non-empty line. However, reasoning models (such as Qwen 2.5 and DeepSeek-R1) standardly format their final answers wrapped in LaTeX boxing (\boxed{...}, $\boxed{...}$), markdown bold (**...**), or inline code tags.

Previously, re.fullmatch(r"[+-]?(?:\d+|\d{1,3}(?:,\d{3})+)", final) on the final line rejected these common wrappers, assigning reward = 0.0 to mathematically correct rollouts and corrupting the RL training reward signal.

Changes

• Unwraps \boxed{...} and strips enclosing markdown/LaTeX delimiters ($, *, `, .) from the final answer line prior to integer matching.
• Restores expected reward 1.0 for \boxed{...} and adds test coverage for $\boxed{...}$, **...**, and `...` in test_train.py.
• Syncs the cookbook documentation snippet in docs/v6/cookbooks/fireworks-rl-training.mdx.

Verification

• Ran test suite via uv run pytest test_train.py -k test_grades_integer_on_last_nonempty_line (22 passed, 100%).
• Passed ruff check and ruff format.

---

Note

Low Risk
Cookbook-only grader and test/doc updates; no auth, infra, or broader SDK behavior changes.

Overview
The Fireworks RL cookbook’s grade_final_integer grader now normalizes common answer wrappers on the last line before parsing the integer, so correct math isn’t scored as wrong when models use typical reasoning-model formatting.

After taking the last non-empty line, it unwraps \boxed{...} and strips enclosing $, *, backticks, whitespace, and trailing periods, then applies the existing comma-aware integer regex. Tests now expect full reward for \boxed{...}, $\boxed{...}$, **...**, and `...` answers. The v6 cookbook doc snippet matches env.py.

^{Reviewed by Cursor Bugbot for commit 3bce83c. Bugbot is set up for automated code reviews on this repo. Configure here.}