pod_prep.sh

#!/bin/bash
# pod_prep.sh — one-shot bootstrap for an RB4 fork pod.
# Prereq: ~/.cache/huggingface/token already written (base64-decoded).
# Installs deps, downloads all scripts + ckpt_280k branch-point + arcmix + tokenizer, arranges data-dir.
set -euo pipefail
export HF_HUB_ENABLE_HF_TRANSFER=0
export HF_TOKEN="$(cat ~/.cache/huggingface/token)"

pip install -q huggingface_hub datasets tokenizers numpy 2>&1 | tail -1
mkdir -p /workspace/run && cd /workspace/run

python - <<'PY'
from huggingface_hub import hf_hub_download as d
R = "SlayerLab/gollem-v5-ckpts"
for f in ["train_gpt_ref.py", "glint_parity_eval.py", "tokenizer.json", "ckpt_autopush.py",
          "fabryka_push.py", "fork_launch.sh", "blend_edu.py", "build_v2_blend.py",
          "build_decontam_index.py", "v1_muon/ckpt_320k.pt", "vals/arcmix_val.bin"]:
    print(d(R, f, local_dir="."))
print(d("SlayerLab/gollem-ui-mirror", "arcmix/train.bin", repo_type="dataset", local_dir="."))
PY

mkdir -p data_arcmix
ln -sf "$(pwd)/arcmix/train.bin" data_arcmix/train.bin
ln -sf "$(pwd)/vals/arcmix_val.bin" data_arcmix/val.bin
echo "=== POD-PREP DONE ==="
echo "arcmix data-dir:"; ls -la data_arcmix/
echo "branch-point:"; ls -la v1_muon/ckpt_320k.pt
echo ""
echo "NEXT — fresh-resume-arcmix (2sigma-seed):"
echo "  mkdir -p run_arcmix_s2 && cp v1_muon/ckpt_320k.pt run_arcmix_s2/ckpt.pt"
echo "  bash fork_launch.sh data_arcmix run_arcmix_s2 forkctrl_arcmix_s2 4242"
echo "NEXT — fork-B (after edu.bin present + blend built):"
echo "  python blend_edu.py --arcmix arcmix/train.bin --edu edu.bin --out data_blend/train.bin --arcmix-tokens 2000000000 --edu-tokens 2000000000"
echo "  cp vals/arcmix_val.bin data_blend/val.bin && mkdir -p run_forkB && cp v1_muon/ckpt_320k.pt run_forkB/ckpt.pt"
echo "  bash fork_launch.sh data_blend run_forkB forkB_arcmix_edu 1337"