NEXUS_discovery_GPU

Backup and experiment package for Intern / discovery GPU sessions (Shanghai A100 lane).

On the Intern machine

/data/NEXUS/                 # JuiceFS ~30G persistent
  logs/                      # smoke, bench, train_runs, heartbeat
  git_backups/work/          # local git (push to this GitHub repo)
  workspace/                 # SESSION_PROGRESS_LATEST.md
  datasets/tiny/

Quick start (on A100 code-server)

cd /data/NEXUS/git_backups/work
# if not cloned yet:
# git clone https://github.com/specimba/NEXUS_discovery_GPU.git /data/NEXUS/git_backups/work

python3 scripts/train_stub.py --steps 100 --size 2048 --seed 42
python3 scripts/gpu_bench.py
bash scripts/tick_1.sh

Agents

• Cline / Kilo Code on code-server: prefer writing proofs to files under /data/NEXUS/logs (shell-integration may be unavailable).
• Explorer often opens /root; open NEXUS / /data/NEXUS for real project files.

GitHub

This repo is the durable backup target. Do not commit PATs or secrets.

Session notes

See docs/ for automation policy (calm single-burst CDP, no thrash).

Session 4 (2026-07-16)

• Longrun inventory + continuity on JuiceFS
• DPO gold + canary config in-repo: datasets/nexus_local/, configs/
• After push, on A100: bash scripts/pull_and_stage.sh
• Laptop OCR helpers under local_ocr/ (optional; not for A100 GPU hours)
• Latest narrative: reports/session4/SESSION_LATEST.md