# EESD stopped checkpoint — 2026-09-21

Repository: https://github.com/yuhanlydia/PBPF
Backup repository: https://huggingface.co/datasets/humanlong/PBPF
Exact source commit: `ed4bdd799816f8a271f3b4cbe122147e79305a90`.

Includes all four `runs/` trees: raw/materialized data, sealed and partial candidate
banks, scores, logs, frozen execution/configuration records and operational scripts.
The final overlay preserves progress made during upload. This is a checkpoint,
not a claim that the experiments or model training are complete.

## Restore on a new inactive machine

Requires git, hf, gzip, tar, sha256sum and Python 3. Use an empty staging directory
and a fresh checkout; these commands must not target a running experiment tree.

```bash
export EESD_STAGE=/root/eesd-migration-stage
mkdir -p "$EESD_STAGE"
hf download humanlong/PBPF --repo-type dataset --local-dir "$EESD_STAGE"
cd "$EESD_STAGE"
sha256sum -c SHA256SUMS
cat PBPF-source.bundle.chunk-* > PBPF-source.bundle
git clone --branch research/eesd-iclr2027 "$EESD_STAGE/PBPF-source.bundle" /root/PBPF
cat eesd-runs-20260921.tar.gz.chunk-* | tar -xzf - -C /root/PBPF
python3 verify_artifacts.py --root /root/PBPF --manifest ARTIFACT_FILES.jsonl
# Replace only the three mutable trees just restored into this fresh checkout.
# Data in runs/eesd-data remains exactly the base archive's data.
rm -rf /root/PBPF/runs/eesd-setup /root/PBPF/runs/eesd-20260920 /root/PBPF/runs/eesd-direct-20260921
cat eesd-final-run-state-20260921.tar.gz.chunk-* | tar -xzf - -C /root/PBPF
python3 verify_artifacts.py --root /root/PBPF --manifest FINAL_ARTIFACT_FILES.jsonl
```

Read `STOP_RECEIPT.json`, `docs/EESD_GPU_MIGRATION_20260921.md`,
`docs/EESD_REMAINING_EXPERIMENTS_20260921.md` and
`configs/experiments/eesd_handoff_tasks_20260921.json` before continuing.
Old PID/adoption/status files describe the old host; `running` in those historical
files does not mean a worker survived shutdown. Partial banks require reconciliation.
The per-file manifest verifies contents, not experiment completeness.

## Models and environment

Base model weights, package caches, `.venv` and credentials are excluded. There
were no trained SFT adapters at checkpoint time. Re-download models/tokenizers at
the recorded revisions, using configs and setup locks; do not substitute latest.
Recreate dependencies from `runs/eesd-setup/requirements.lock.txt` and its environment
inventory. Existing runtime locks bind the old host/binaries: preserve them as
history and create a new execution profile/amendment for the replacement machine.
The current scheduling scripts do not automatically schedule four GPUs.

Evaluator/private views contain benchmark labels and must remain outside model
prompts/generation inputs. Future experiment tasks include unimplemented direct
Replay adapter work; this backup is not a promise of one-command execution.