Initial K3 snapshot: 0.5B KDA/MLA/MoE train path

Standalone tree split from LLMRL/projects/kda. Includes Triton dt_bias
backward fix, train_k3 --preset 0.5b, SFT, Docker runtime, and tests.

scripts/container_help.py

@@ -0,0 +1,43 @@
+"""Default container command: print usage and refuse to start a silent toy run."""
+
+from __future__ import annotations
+
+HELP = """kda image: one GPU runtime for train / monitor / eval.
+
+Do not docker run without a command — toy overfit is not the default.
+
+Volumes (host → container):
+  ./ckpts            → /workspace/kda/ckpts
+  pretrain corpus    → /data/pretrain   (ro)
+  SFT bitext         → /data/sft        (ro)
+  frozen eval        → /data/eval       (ro)
+  HF tokenizer cache → /cache/huggingface
+  SWANLAB_API_KEY    via -e  (never COPY into the image)
+
+Examples:
+  docker run --rm --gpus all kda:<tag> python train.py
+  docker run --rm --gpus all -e SWANLAB_API_KEY \\
+    -v "$PWD/ckpts:/workspace/kda/ckpts" \\
+    kda:<tag> python train_k3.py --preset toy --attnres off
+  docker run --rm --gpus all kda:<tag> swanlab ping
+  docker run --rm --gpus all \\
+    -v "$PWD/ckpts:/workspace/kda/ckpts" -v "$PWD/data/eval:/data/eval:ro" \\
+    kda:<tag> python -m kda.training.eval_mt \\
+      --ckpt /workspace/kda/ckpts/k3_wiki.pt \\
+      --src /data/eval/zh2en.src.txt --ref /data/eval/zh2en.ref.txt \\
+      --target-lang en
+  docker run --rm --gpus all \\
+    -v "$PWD/ckpts:/workspace/kda/ckpts" -v "$PWD/data/sft:/data/sft:ro" \\
+    kda:<tag> python train_sft.py --ckpt /workspace/kda/ckpts/k3_wiki.pt \\
+      --data /data/sft/toy.jsonl
+  docker run --rm --gpus all kda:<tag> python -m pytest -q
+"""
+
+
+def main() -> None:
+    print(HELP)
+    raise SystemExit(2)
+
+
+if __name__ == "__main__":
+    main()

scripts/export_flores.py

@@ -0,0 +1,61 @@
+"""Export FLORES-200 zh↔en into line-aligned src/ref files for eval_mt.
+
+Corpus is not committed. Typical:
+
+    uv run python scripts/export_flores.py --out data/eval
+"""
+
+from __future__ import annotations
+
+import argparse
+from pathlib import Path
+
+
+def _write(path: Path, lines: list[str]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text("\n".join(lines) + "\n", encoding="utf-8")
+
+
+def main() -> None:
+    p = argparse.ArgumentParser(description=__doc__)
+    p.add_argument("--out", default="data/eval")
+    p.add_argument("--split", default="devtest", choices=["dev", "devtest"])
+    args = p.parse_args()
+
+    from datasets import load_dataset
+
+    ds = None
+    err: Exception | None = None
+    for config in ("eng_Latn-zho_Hans", "default"):
+        try:
+            ds = load_dataset("facebook/flores", config, split=args.split)
+            break
+        except Exception as exc:  # noqa: BLE001 — try the next config name
+            err = exc
+    if ds is None:
+        raise SystemExit(f"could not load facebook/flores ({err})")
+
+    cols = set(ds.column_names)
+    en_key = next(
+        (c for c in ("sentence_eng_Latn", "eng_Latn", "sentence_en") if c in cols),
+        None,
+    )
+    zh_key = next(
+        (c for c in ("sentence_zho_Hans", "zho_Hans", "sentence_zh") if c in cols),
+        None,
+    )
+    if en_key is None or zh_key is None:
+        raise SystemExit(f"FLORES columns not found: {sorted(cols)}")
+
+    en = [row[en_key].strip() for row in ds]
+    zh = [row[zh_key].strip() for row in ds]
+    out = Path(args.out)
+    _write(out / "flores.zh2en.src.txt", zh)
+    _write(out / "flores.zh2en.ref.txt", en)
+    _write(out / "flores.en2zh.src.txt", en)
+    _write(out / "flores.en2zh.ref.txt", zh)
+    print(f"wrote {len(zh)} pairs under {out}/flores.*.txt")
+
+
+if __name__ == "__main__":
+    main()