Pull OPUS-100 en-zh for SFT instead of a checked-in jsonl

train_sft --data opus-100 streams Helsinki-NLP/opus-100, writes both
directions, and skips frozen eval sentences. Runtime cache stays under
data/sft/ (gitignored).

tests/integration/test_sft_data.py

@@ -1,4 +1,11 @@
-from kda.training.data import IGNORE_INDEX, collate_sft, encode_sft_row, load_sft_rows
+from kda.training.data import (
+    IGNORE_INDEX,
+    collate_sft,
+    encode_sft_row,
+    fetch_opus100_enzh,
+    load_sft_rows,
+    resolve_sft_rows,
+)
 from kda.training.prompts import instruction_prompt
 
 
@@ -42,6 +49,36 @@ def test_collate_and_jsonl(tmp_path):
     assert (y == IGNORE_INDEX).any()
 
 
+def test_resolve_sft_rows_reads_local_jsonl(tmp_path):
+    path = tmp_path / "bitext.jsonl"
+    path.write_text(
+        '{"src": "你好", "tgt": "Hello", "target_lang": "en"}\n',
+        encoding="utf-8",
+    )
+    rows = resolve_sft_rows(str(path))
+    assert rows == [{"src": "你好", "tgt": "Hello", "target_lang": "en"}]
+
+
+def test_fetch_opus_skips_eval_sentences(tmp_path, monkeypatch):
+    eval_dir = tmp_path / "eval"
+    eval_dir.mkdir()
+    (eval_dir / "zh2en.src.txt").write_text("禁止句\n", encoding="utf-8")
+
+    class _DS:
+        def __iter__(self):
+            yield {"translation": {"en": "Hello", "zh": "你好"}}
+            yield {"translation": {"en": "skip", "zh": "禁止句"}}
+            yield {"translation": {"en": "Thanks", "zh": "谢谢"}}
+
+    fake = type("datasets", (), {"load_dataset": staticmethod(lambda *a, **k: _DS())})
+    monkeypatch.setitem(__import__("sys").modules, "datasets", fake)
+    rows = fetch_opus100_enzh(10, cache_dir=tmp_path / "sft", eval_dir=eval_dir)
+    srcs = {r["src"] for r in rows}
+    assert "禁止句" not in srcs
+    assert "你好" in srcs and "Hello" in srcs
+    assert "谢谢" in srcs and "Thanks" in srcs
+
+
 def test_toy_sft_file_parses():
     from pathlib import Path