Unify parquet schema across sft/pretrain splits

data/pretrain.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9a3e9223571324cd846af0dd8681e035dc1516ee3fcff76e2d8f665e03669af3
-size 79001
+oid sha256:5807e85385d98135a3ae7691f7c87032d6ad3a50bbc75123bfda04afdca3d55f
+size 79158

data/sft.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d04bd1045c118250c39185a030adcbc374aafce16e6b27e76971cc8ea57958e5
-size 221367
+oid sha256:0e189ea81c104069543b726422ec38621b06598299059475c65277ac05c176ee
+size 218841

train_and_upload.py

@@ -0,0 +1,132 @@
+"""Fine-tune and publish a Japanese data engineering mentor on the DE2026 dataset."""
+
+from __future__ import annotations
+
+import os
+import shutil
+from pathlib import Path
+
+from dotenv import load_dotenv
+
+ROOT = Path(__file__).resolve().parents[3]
+load_dotenv(ROOT / ".env")
+
+import torch
+from datasets import load_dataset
+from huggingface_hub import HfApi
+from peft import LoraConfig, get_peft_model
+from transformers import (
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    DataCollatorForLanguageModeling,
+    Trainer,
+    TrainingArguments,
+)
+
+BASE_MODEL = "Qwen/Qwen2.5-0.5B-Instruct"
+DATASET_ID = "LoNebula/data-engineer-2026-ja"
+REPO_ID = "LoNebula/data-engineer-2026-ja-mentor"
+OUTPUT_DIR = ROOT / "dev" / "hf-de2026-dataset" / "model"
+MAX_LENGTH = 1024
+
+
+def build_messages(row: dict[str, object]) -> list[dict[str, str]]:
+    """Render an SFT row into chat messages; pass pretrain text through unchanged."""
+    if "output" in row:
+        instruction = str(row.get("instruction") or "").strip()
+        context = str(row.get("input") or "").strip()
+        user_content = instruction if not context else f"{instruction}\n\n{context}"
+        return [
+            {
+                "role": "system",
+                "content": (
+                    "あなたはデータエンジニアringのメンターです。2026年時点の実務知識を、"
+                    "具体的で検証可能な手順/detailsを交えて日本語で説明してください。"
+                    "version-sensitive なアドバイスには公式ドキュメントを参照するよう指示してください。"
+                ),
+            },
+            {"role": "user", "content": user_content},
+            {"role": "assistant", "content": str(row["output"])},
+        ]
+    return [{"role": "user", "content": str(row["text"])}]
+
+
+def main() -> None:
+    token = os.environ.get("HF_TOKEN")
+    if not token:
+        raise RuntimeError("HF_TOKEN is missing from the project environment")
+    api = HfApi(token=token)
+    print(f"Hugging Face account: {api.whoami()['name']}")
+
+    dataset = load_dataset(DATASET_ID, token=token)
+    tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, token=token)
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
+
+    def tokenize(row: dict[str, object]) -> dict[str, list[int]]:
+        messages = build_messages(row)
+        text = tokenizer.apply_chat_template(
+            messages, tokenize=False, add_generation_prompt=False
+        )
+        return tokenizer(text, truncation=True, max_length=MAX_LENGTH)
+
+    prepared = dataset.map(tokenize, remove_columns=dataset["sft"].column_names)
+    dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+    model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, token=token, torch_dtype=dtype)
+    model = get_peft_model(
+        model,
+        LoraConfig(
+            r=8,
+            lora_alpha=16,
+            lora_dropout=0.05,
+            target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
+            task_type="CAUSAL_LM",
+        ),
+    )
+    OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
+    args = TrainingArguments(
+        output_dir=str(OUTPUT_DIR / "checkpoints"),
+        num_train_epochs=3,
+        per_device_train_batch_size=1,
+        per_device_eval_batch_size=1,
+        gradient_accumulation_steps=4,
+        learning_rate=1e-4,
+        logging_steps=10,
+        eval_strategy="epoch",
+        save_strategy="no",
+        report_to="none",
+        fp16=torch.cuda.is_available(),
+        optim="adamw_torch",
+    )
+    trainer = Trainer(
+        model=model,
+        args=args,
+        train_dataset=prepared["sft"],
+        eval_dataset=prepared["sft"].select(range(min(64, len(prepared["sft"])))),
+        data_collator=DataCollatorForLanguageModeling(tokenizer, mlm=False),
+    )
+    trainer.train()
+    print(f"Held-out loss: {trainer.evaluate()['eval_loss']:.4f}")
+
+    merged = model.merge_and_unload()
+    merged.save_pretrained(OUTPUT_DIR, safe_serialization=True)
+    tokenizer.save_pretrained(OUTPUT_DIR)
+    shutil.copy2(
+        ROOT / "dev" / "hf-de2026-dataset" / "README.md",
+        OUTPUT_DIR / "README.md",
+    )
+    api.create_repo(REPO_ID, repo_type="model", exist_ok=True, private=False)
+    api.upload_folder(
+        folder_path=str(OUTPUT_DIR),
+        repo_id=REPO_ID,
+        repo_type="model",
+        commit_message="Publish Japanese data engineering mentor fine-tune on DE2026",
+    )
+    print(f"Published model: https://huggingface.co/{REPO_ID}")
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except Exception as exc:
+        raise SystemExit(f"Model training or upload failed ({type(exc).__name__}); details suppressed.") from None