prepare_data.py

"""Phase 1: build the tokenizer and the pretokenized corpus, upload to HF.

Steps:
  1. stream the source dataset (hy Wikipedia by default) and dump a text sample
  2. train a SentencePiece unigram tokenizer on that sample
  3. tokenize the FULL corpus into a flat uint16 stream -> train.bin / val.bin
  4. upload tokenizer + .bin files to the HF dataset repo

Run this on a high-RAM / many-CPU Colab runtime (or a strong local machine).
It does NOT need a TPU. Then run launch.py for the TPU training.
"""
from __future__ import annotations

import os
import time

import numpy as np


def _log(m: str) -> None:
    print(f"[prepare {time.strftime('%H:%M:%S')}] {m}", flush=True)


# Heartbeat so the supervisor sees prepare is progressing across its long,
# blocking phases (SentencePiece training, corpus tokenization). run_all injects
# the real writer; default is a no-op when prepare_data is run standalone.
def _default_beat(_m: str) -> None:
    from pathlib import Path
    import time as _t
    try:
        p = Path("/content/train_logs/heartbeat.txt")
        p.parent.mkdir(parents=True, exist_ok=True)
        p.write_text(f"{_t.time():.0f} {_m}\n", encoding="utf-8")
    except Exception:
        pass


def main(beat=_default_beat) -> None:
    import sentencepiece as spm
    from datasets import load_dataset
    from huggingface_hub import HfApi

    from config import (MODEL, TOKENIZER, DATA_REPO, SOURCE_DATASET,
                        SOURCE_CONFIG, SOURCE_SPLIT)

    hf_token = os.environ["HF_TOKEN"]
    api = HfApi(token=hf_token)

    def text_column(ds) -> str:
        for c in ("text", "content", "raw_content", "document"):
            if c in ds.column_names:
                return c
        return ds.column_names[0]

    # --- 1. sample text for tokenizer training ---
    _log(f"loading {SOURCE_DATASET}:{SOURCE_CONFIG} (streaming)")
    beat("prepare: sampling text")
    stream = load_dataset(SOURCE_DATASET, SOURCE_CONFIG, split=SOURCE_SPLIT,
                          streaming=True, token=hf_token)
    col = text_column(stream)
    _log(f"text column: {col}")

    sample_path = "tok_sample.txt"
    n = 0
    with open(sample_path, "w", encoding="utf-8") as f:
        for row in stream:
            t = (row.get(col) or "").strip()
            if not t:
                continue
            f.write(t[: TOKENIZER.max_chars_per_row].replace("\n", " ") + "\n")
            n += 1
            if n >= TOKENIZER.train_sample_rows:
                break
    _log(f"wrote {n:,} rows for tokenizer training")

    # --- 2. train SentencePiece ---
    _log("training SentencePiece tokenizer")
    beat("prepare: training tokenizer")
    spm.SentencePieceTrainer.train(
        input=sample_path,
        model_prefix="armenian_sp",
        vocab_size=TOKENIZER.vocab_size,
        model_type=TOKENIZER.model_type,
        character_coverage=TOKENIZER.character_coverage,
        input_sentence_size=n,
        shuffle_input_sentence=True,
        bos_id=1, eos_id=2, unk_id=0, pad_id=3,
        num_threads=os.cpu_count() or 8,
    )
    sp = spm.SentencePieceProcessor(model_file="armenian_sp.model")
    assert sp.vocab_size() == MODEL.vocab_size, (
        f"vocab mismatch: tokenizer {sp.vocab_size()} vs model {MODEL.vocab_size}")

    # --- 3. tokenize full corpus in parallel across all host cores ---
    n_proc = max(1, (os.cpu_count() or 8))
    _log(f"tokenizing full corpus with num_proc={n_proc}")
    beat(f"prepare: tokenizing corpus ({n_proc} cores)")
    # Non-streaming so datasets.map can shard across processes (hy wiki is small).
    full = load_dataset(SOURCE_DATASET, SOURCE_CONFIG, split=SOURCE_SPLIT, token=hf_token)

    sp_model_path = os.path.abspath("armenian_sp.model")

    def tok_batch(batch):
        # Each worker builds its own processor (SentencePieceProcessor isn't picklable).
        proc = tok_batch._sp
        if proc is None:
            proc = spm.SentencePieceProcessor(model_file=sp_model_path)
            tok_batch._sp = proc
        eos_id = proc.eos_id()
        out = []
        for t in batch[col]:
            t = (t or "").strip()
            if not t:
                out.append([])
                continue
            ids = proc.encode(t, out_type=int)
            ids.append(eos_id)
            out.append(ids)
        return {"ids": out}
    tok_batch._sp = None

    tokenized = full.map(tok_batch, batched=True, batch_size=1000,
                         num_proc=n_proc, remove_columns=full.column_names,
                         desc="tokenize")

    # Concatenate all id lists into one flat uint16 stream.
    _log("concatenating token stream")
    parts = [np.asarray(x, dtype=np.uint16) for x in tokenized["ids"] if x]
    all_ids = np.concatenate(parts)
    _log(f"total tokens: {len(all_ids):,}")

    # --- 4. split + write .bin ---
    n_val = max(1, int(len(all_ids) * 0.005))
    all_ids[:-n_val].tofile("train.bin")
    all_ids[-n_val:].tofile("val.bin")
    _log(f"train.bin {len(all_ids) - n_val:,} | val.bin {n_val:,}")

    # --- 5. upload ---
    api.create_repo(DATA_REPO, repo_type="dataset", exist_ok=True)
    for fn in ("train.bin", "val.bin", "armenian_sp.model", "armenian_sp.vocab"):
        _log(f"uploading {fn}")
        api.upload_file(path_or_fileobj=fn, path_in_repo=fn,
                        repo_id=DATA_REPO, repo_type="dataset")
    _log("done. data ready on HF.")


if __name__ == "__main__":
    main()