Fix get_nl_ratio ZeroDivisionError on empty document text

[Chessing234]

Bug

python/dolma/taggers/code/starcoder.py::get_nl_ratio computes the comment-to-code ratio as:

def get_nl_ratio(text, language):
    """get the ratio of comments to code in a program"""
    if language == "python":
        comments = get_text_python(text)
        ratio = len(comments) / len(text)
    else:
        ratio = comment_size(text, language) / len(text)
    return ratio

When text is empty (0 bytes), both branches divide by 0 and raise ZeroDivisionError.

This is reachable from CodeStarCoderTaggers2.predict in python/dolma/taggers/code/code_taggers.py:253, which calls get_nl_ratio(doc.text, lang) without a try / except when the detected language is Python/Java/JavaScript. An empty-source document (e.g. a 0-byte file that made it through earlier filters, or a document whose body got stripped upstream) therefore aborts the tagger mid-run and takes the current shard's processing with it.

The older CodeStarCoderTaggers (v1, code_taggers.py:209) does wrap the call in try: ... except: nl_ratio = -1.0, so v1 silently coerces the crash into a sentinel; v2 has no such safety net.

Fix

Short-circuit in get_nl_ratio for empty text — before any division — returning 0.0:

 def get_nl_ratio(text, language):
     """get the ratio of comments to code in a program"""
+    if not text:
+        return 0.0
     if language == "python":
         comments = get_text_python(text)
         ratio = len(comments) / len(text)
     else:
         ratio = comment_size(text, language) / len(text)
     return ratio

• Empty text → 0.0: natural "zero comments over zero code" answer, keeps the tagger's output schema consistent — downstream filters that check nl_ratio_doc / code_to_comment_ratio spans see a well-defined float instead of a missing / erroring span or v1's -1.0 magic number.
• Non-empty text: unchanged — both language branches execute exactly as before.

Two added lines, no other changes.