hardneg_2.py

import os
from pathlib import Path
import torch
import pandas as pd
from datasets import load_dataset, Dataset, load_from_disk
from sentence_transformers import (
    SentenceTransformer,
    SentenceTransformerTrainer,
    SentenceTransformerTrainingArguments,
)
from sentence_transformers.losses import CachedMultipleNegativesRankingLoss
from sentence_transformers.training_args import BatchSamplers
from sentence_transformers.util import mine_hard_negatives


# =========================
# CONFIG
# =========================
DATASET_NAME = "phamson02/large-vi-legal-queries"

# Stage 1 model dir
BASE_MODEL_DIR = "./embeddinggemma-300m-vilegal"
BASE_MODEL_CHECKPOINT_DIR = "./embeddinggemma-300m-vilegal"

# Stage 2 output dir
STAGE2_OUTPUT_DIR = "./embeddinggemma-300m-vilegal-stage2-hardneg"

# Cache dir
CACHE_ROOT = Path("./cache_vilegal_stage2")
CACHE_ROOT.mkdir(parents=True, exist_ok=True)

# Cached artifacts
CLEAN_DF_PATH = CACHE_ROOT / "clean_df.parquet"
PAIR_DATASET_PATH = CACHE_ROOT / "pair_dataset"
HARD_NEGATIVE_DATA_DIR = CACHE_ROOT / "hardneg_dataset"

TASK_NAME = "Retrieval"

# Data control
LIMIT_ROWS = None  # set e.g. 20000 for quick tests

# Cache toggles
USE_CACHE_CLEAN_DF = True
USE_CACHE_PAIR_DATASET = True
USE_CACHE_HARDNEG = True
FORCE_REBUILD_CLEAN_DF = False
FORCE_REBUILD_PAIR_DATASET = False
FORCE_REBUILD_HARDNEG = False

# Training control
RESUME_IF_POSSIBLE = True


# =========================
# HELPERS
# =========================
def clean_text(x):
    if x is None:
        return ""
    x = str(x).strip()
    x = " ".join(x.split())
    return x


def build_positive_document(row):
    return f"text: {row['context']}"


def path_exists_and_nonempty(path: Path) -> bool:
    return path.exists() and any(path.iterdir()) if path.is_dir() else path.exists()


def get_last_checkpoint(output_dir: str):
    output_path = Path(output_dir)
    if not output_path.exists():
        return None

    checkpoints = []
    for p in output_path.iterdir():
        if p.is_dir() and p.name.startswith("checkpoint-"):
            try:
                step = int(p.name.split("-")[-1])
                checkpoints.append((step, p))
            except ValueError:
                continue

    if not checkpoints:
        return None

    checkpoints.sort(key=lambda x: x[0])
    return str(checkpoints[-1][1])


# =========================
# DATA PREP
# =========================
def load_and_prepare_dataframe(limit_rows=None):
    if (
        USE_CACHE_CLEAN_DF
        and not FORCE_REBUILD_CLEAN_DF
        and CLEAN_DF_PATH.exists()
    ):
        print(f"✅ Loading cached clean dataframe from: {CLEAN_DF_PATH}")
        df = pd.read_parquet(CLEAN_DF_PATH)
        print("Cached clean rows:", len(df))
        return df

    print("📥 Loading raw dataset from hub...")
    ds = load_dataset(DATASET_NAME, split="train")

    if limit_rows is not None:
        ds = ds.select(range(min(limit_rows, len(ds))))

    df = ds.to_pandas()
    print("Raw shape:", df.shape)

    for col in ["domain", "title", "header", "aspect", "context", "query"]:
        if col not in df.columns:
            df[col] = ""
        df[col] = df[col].apply(clean_text)

    df = df[(df["query"] != "") & (df["context"] != "")]
    df = df.drop_duplicates(subset=["query", "context"]).reset_index(drop=True)

    print("Cleaned rows:", len(df))

    if USE_CACHE_CLEAN_DF:
        print(f"💾 Saving clean dataframe cache to: {CLEAN_DF_PATH}")
        df.to_parquet(CLEAN_DF_PATH, index=False)

    return df


def build_pair_dataset(df):
    if (
        USE_CACHE_PAIR_DATASET
        and not FORCE_REBUILD_PAIR_DATASET
        and path_exists_and_nonempty(PAIR_DATASET_PATH)
    ):
        print(f"✅ Loading cached pair dataset from: {PAIR_DATASET_PATH}")
        dataset = load_from_disk(str(PAIR_DATASET_PATH))
        print("Cached pair dataset:", dataset)
        return dataset

    print("🛠 Building pair dataset...")
    pair_df = pd.DataFrame(
        {
            "query": df["query"].tolist(),
            "positive": df.apply(build_positive_document, axis=1).tolist(),
        }
    )

    dataset = Dataset.from_pandas(pair_df, preserve_index=False)

    if USE_CACHE_PAIR_DATASET:
        print(f"💾 Saving pair dataset cache to: {PAIR_DATASET_PATH}")
        dataset.save_to_disk(str(PAIR_DATASET_PATH))

    return dataset


# =========================
# HARD NEGATIVE MINING
# =========================
def mine_hard_negative_dataset(pair_dataset, model_dir):
    if (
        USE_CACHE_HARDNEG
        and not FORCE_REBUILD_HARDNEG
        and path_exists_and_nonempty(HARD_NEGATIVE_DATA_DIR)
    ):
        print(f"✅ Loading cached hard negative dataset from: {HARD_NEGATIVE_DATA_DIR}")
        hn_dataset = load_from_disk(str(HARD_NEGATIVE_DATA_DIR))
        print("Cached hard negative dataset:", hn_dataset)
        return hn_dataset

    print("⛏ Mining hard negatives...")
    miner_model = SentenceTransformer(model_dir)
    miner_model.max_seq_length = 512

    hn_dataset = mine_hard_negatives(
        dataset=pair_dataset,
        model=miner_model,
        positive_column_name="positive",
        range_min=10,
        range_max=50,
        relative_margin=0.05,
        num_negatives=3,
        sampling_strategy="random",
        batch_size=128,
        use_faiss=True,
        query_prompt_name="query",
        corpus_prompt_name="document",
        output_format="n-tuple",
        use_multi_process=True,
    )

    if USE_CACHE_HARDNEG:
        print(f"💾 Saving hard negative dataset cache to: {HARD_NEGATIVE_DATA_DIR}")
        hn_dataset.save_to_disk(str(HARD_NEGATIVE_DATA_DIR))

    return hn_dataset


def preview_hard_negatives(hn_dataset, sample_size=10):
    if len(hn_dataset) == 0:
        print("No hard negatives to preview.")
        return

    sample_df = hn_dataset.to_pandas().sample(
        min(sample_size, len(hn_dataset)),
        random_state=42,
    )

    for _, row in sample_df.iterrows():
        print("=" * 100)
        print("QUERY:\n", row["query"])
        print("\nPOSITIVE:\n", row["positive"][:700])

        for i in range(1, 10):
            neg_col = f"negative_{i}"
            if neg_col in row and isinstance(row[neg_col], str):
                print(f"\n{neg_col.upper()}:\n", row[neg_col][:700])

        print()


# =========================
# TRAINING
# =========================
def train_stage2_with_hardneg(hn_dataset, model_checkpoint_dir, output_dir):
    # IMPORTANT: không dùng .to("cuda") khi chạy torchrun
    train_model = SentenceTransformer(model_checkpoint_dir)
    train_model.max_seq_length = 512

    loss = CachedMultipleNegativesRankingLoss(
        train_model,
        mini_batch_size=32,
        gather_across_devices=True,
    )

    training_args = SentenceTransformerTrainingArguments(
        prompts=train_model.prompts[TASK_NAME],
        torch_compile=False,
        output_dir=output_dir,
        num_train_epochs=1,
        per_device_train_batch_size=1024,
        gradient_accumulation_steps=1,
        learning_rate=1e-5,
        warmup_ratio=0.1,
        bf16=torch.cuda.is_available(),
        logging_steps=50,
        save_strategy="epoch",
        report_to="none",
        remove_unused_columns=False,
        batch_sampler=BatchSamplers.NO_DUPLICATES,
        dataloader_num_workers=8,
        dataloader_persistent_workers=True,
        dataloader_drop_last=True,
        ddp_find_unused_parameters=False,
    )

    trainer = SentenceTransformerTrainer(
        model=train_model,
        args=training_args,
        train_dataset=hn_dataset,
        loss=loss,
    )

    resume_checkpoint = None
    if RESUME_IF_POSSIBLE:
        resume_checkpoint = get_last_checkpoint(output_dir)
        if resume_checkpoint is not None:
            print(f"🔁 Resuming from checkpoint: {resume_checkpoint}")
        else:
            print("ℹ️ No checkpoint found. Training from scratch.")

    trainer.train(resume_from_checkpoint=resume_checkpoint)

    print(f"💾 Saving final model to: {output_dir}")
    trainer.save_model(output_dir)


# =========================
# MAIN
# =========================
def main():
    # 1) Load + clean
    df = load_and_prepare_dataframe(limit_rows=LIMIT_ROWS)

    # 2) Build query-positive dataset
    pair_dataset = build_pair_dataset(df)
    print("Pair dataset:", pair_dataset)
    if len(pair_dataset) > 0:
        print("Sample pair:", pair_dataset[0])

    # 3) Mine hard negatives từ stage 1 model
    hn_dataset = mine_hard_negative_dataset(
        pair_dataset=pair_dataset,
        model_dir=BASE_MODEL_DIR,
    )
    print("Hard negative dataset:", hn_dataset)
    if len(hn_dataset) > 0:
        print("Sample n-tuple:", hn_dataset[0])

    # 4) Preview vài sample hard negative
    preview_hard_negatives(hn_dataset, sample_size=10)

    # 5) Train stage 2 từ checkpoint stage 1
    train_stage2_with_hardneg(
        hn_dataset=hn_dataset,
        model_checkpoint_dir=BASE_MODEL_CHECKPOINT_DIR,
        output_dir=STAGE2_OUTPUT_DIR,
    )


if __name__ == "__main__":
    main()