amy-checkpoints / README.md
Andandd's picture
Add README with decryption instructions
537b1bb verified
|
Raw History Blame Contribute Delete
1.52 kB

Amy Checkpoints (Encrypted)

All .pt checkpoint files are encrypted with AES-256-GCM.

Structure

  • v1_cache_encrypted/ β€” Best models, V2 steps, EMA variants, SWA, training milestones (22 files)
  • v1_desktop_encrypted/ β€” Full training run series from Desktop (174 files)
  • decrypt_checkpoints.py β€” Decryption script

How to decrypt

  1. Install dependencies:

    pip install cryptography huggingface_hub
    
  2. Clone this repo:

    huggingface-cli download Andandd/amy-checkpoints --local-dir ./amy-checkpoints
    
  3. Set your decryption key and run:

    export AMY_DECRYPT_KEY=<your-key-here>
    python decrypt_checkpoints.py ./amy-checkpoints/v1_cache_encrypted ./checkpoints_decrypted
    

Encryption details

  • Algorithm: AES-256-GCM (authenticated encryption)
  • Each file: [12-byte random nonce][ciphertext + 16-byte auth tag]
  • Key is NOT stored in this repo

File inventory

v1_cache (C: drive, curated)

  • amy_best.pt, amy_best_ema.pt β€” Best model + exponential moving average
  • amy_best_old_507M.pt, amy_best_ema_old_507M.pt β€” Earlier 507M architecture
  • amy_swa.pt β€” Stochastic weight average
  • amy_v2_step_*.pt β€” V2 architecture training (steps 100, 200, 1000)
  • amy_train_*/amy_emergency_* β€” Training run milestones (May 2026)
  • amy_curriculum.pt, amy_spaced_rep.pt, amy_learnability_rw.pt, amy_selective_loss.pt β€” Training metadata

v1_desktop (full series)

  • 174 checkpoints from April 2026 training runs (steps 1 through 16000)