CCCCCyx/20260526_repro_pkg

repro_pkg — portable trajectory reproducer

Self-contained package to compute the same (iter, cos_eff, E_phi, E_perp, p_phi) trajectory as report_full2/data/trajectory_dir_energy.csv, for a new optimizer's HF checkpoint sequence, on any machine with python + torch.

What's inside

repro_pkg/
  reproduce.py                  standalone trajectory script (no project deps)
  build_pkg.py                  how this package was built (kept for reference)
  example_manifest.json         template — fill in your own ckpts
  phi_hat/                      6 × .npz, shared-direction unit vectors (~328 MB)
  theta_0/                      6 × .npz, Qwen3-8B-Base for the 6 params (~440 MB)
  README.md                     this file

Total: ~770 MB. No need to ship the 16 GB base model or the 2.9 GB full phi_hat cache.

Why only 6 parameters?

The trajectory tables and figA/figB aggregate sum_p over 6 representative parameters, not all 41 LM params. The 6 cover MLP down-proj across depth (layers 7/14/21/28) and one attention block (layer 21 q+o). n_params=6 in every row of the output CSV.

(The terminal-state table data/energy_per_opt.csv does use the full 34 params — if you need that, repackage with PHI_DIR pointing at the full data/weight/energy/phi_hat/ and update PARAMS to the union.)

How to reproduce on another machine

1. Copy the whole repro_pkg/ directory

rsync -ah repro_pkg/  user@host:/dst/path/

2. Install runtime

pip install numpy pandas torch safetensors

3. Produce HF checkpoints for your optimizer

You already have this — a directory per training iter, each containing model-*.safetensors + model.safetensors.index.json, in Qwen3-8B-Base layout (same parameter names).

4. Write your own manifest

Copy example_manifest.json to my_manifest.json and fill in:

{
  "optimizers": {
    "myopt": {
      "checkpoints": [
        {"iter": 100, "hf_path": "/abs/path/to/myopt/iter_0000100-hf"},
        {"iter": 200, "hf_path": "/abs/path/to/myopt/iter_0000200-hf"},
        {"iter": 300, "hf_path": "/abs/path/to/myopt/iter_0000300-hf"}
      ]
    }
  }
}

You can list one or many optimizers; pick one with --opt myopt.

5. Run

python reproduce.py --manifest my_manifest.json --opt myopt --out traj_myopt.csv

Output CSV columns (identical to source-of-truth report_full2/data/trajectory_dir_energy.csv):

opt, iter, n_params, E_tau, E_phi, E_perp, p_phi, p_perp, cos_eff

Methodology recap

For each checkpoint at training step t:

tau_t,p      = theta_t,p - theta_0,p                   per representative param p
alpha_p(t)   = <tau_t,p , phi_hat_p>                   projection onto shared dir
E_phi(t)     = sum_p alpha_p(t)^2                      along phi*
E_tau(t)     = sum_p ||tau_t,p||_F^2                   total accumulated energy
E_perp(t)    = E_tau(t) - E_phi(t)                     residual / off-axis
p_phi(t)     = E_phi(t) / E_tau(t)                     fraction on shared dir
cos_eff(t)   = sqrt(p_phi(t))                          energy-equivalent cos

phi_hat = unit(tau_adam_final), cached. The 7 healthy optimizers in the original study agree on this direction to within 1% (cos ≥ 0.989 to Adam), so it is the canonical "shared RL direction".

Interpretation cheatsheet (from report_full2/review5_trajectory_explanation.md)

pattern	meaning
cos_eff ~ 1.0, E_tau flat	healthy — energy concentrated on phi*
cos_eff ~ 1.0, E_tau flat, low reward	weak — accumulates correctly but RL no-op (e.g. SOAP)
cos_eff ~ 1.0, sudden per-step grad spike	step-energy collapse (e.g. Lion @ 80)
cos_eff monotonic decline, E_perp >> E_phi	diffuse / off-axis (e.g. SGD, RMS)

Sanity check (optional)

To verify the package reproduces the original numbers, point the manifest at the Adam HF checkpoints from this project and compare:

opt=adam iter=31    cos_eff=0.9994  E_tau=138.65  p_phi=0.9987
opt=adam iter=1023  cos_eff=1.0000  E_tau=138.84  p_phi=1.0000

These should match report_full2/data/trajectory_dir_energy.csv to all shown digits.