#!/bin/bash
# Tokenize FineWeb with the pre-trained 16K BPE tokenizer.
# The tokenizer model is committed to the repo; this script only needs to
# download the raw docs and tokenize them into binary shards.
# Run from experiments/quintary/ on the pod.
set -e

# Ensure conda env is active
source ~/miniconda3/etc/profile.d/conda.sh 2>/dev/null || true
conda activate golf 2>/dev/null || true

echo "=== Building 16K dataset (reusing pre-trained tokenizer) ==="
echo ""

# Clone upstream if not present (gitignored in our repo)
UPSTREAM_DIR="../../upstream"
if [ ! -d "$UPSTREAM_DIR/data" ]; then
    echo "Cloning openai/parameter-golf..."
    git clone --depth 1 https://github.com/openai/parameter-golf.git "$UPSTREAM_DIR"
fi

# Reuse the committed tokenizer model — skips SentencePiece training
python3 "$UPSTREAM_DIR/data/download_hf_docs_and_tokenize.py" \
  --tokenizer-config ./tokenizer_specs_16k.json \
  --output-root ./data \
  --skip-byte \
  --reuse-sp-model 16384=./fineweb_16384_bpe.model

echo ""
echo "=== Done ==="
echo "Tokenizer: ./data/tokenizers/fineweb_16384_bpe.model"
echo "Dataset:   ./data/datasets/fineweb10B_sp16384/"
echo ""
echo "Run training with:"
echo "  VOCAB_SIZE=16384 TOKENIZER_PATH=./data/tokenizers/fineweb_16384_bpe.model \\"
echo "  DATA_PATH=./data/datasets/fineweb10B_sp16384 \\"
echo "  bash run_cuda_quintary.sh"