PersonaCore Milestone 1 — slim inference checkpoint

Slim inference checkpoint for PersonaCore Milestone 1 — a from-scratch 13.9M-parameter GPT trained on-device (Apple Silicon, fp32/MPS).

Artifact

model_slim.pt (~55.6 MB) — inference weights only. The full training checkpoint (optimizer/scheduler/RNG state) is not published.

Provenance (embedded in the file itself)

• Training commit: 3a46815 — the artifact carries its own git SHA and ModelConfig, so the exact architecture and code version reconstruct from the file with no out-of-band knowledge
• Step: 49,000 (best-validation checkpoint of the 50,000-step run)
• Parameters: 13,891,584 (tied embedding counted once)
• Headline: deterministic full-validation perplexity 2.1066 over 12,636,922 scored target tokens

Loading — weights_only=True is mandatory

import torch
ckpt = torch.load("checkpoints/model_slim.pt", map_location="cpu", weights_only=True)

The file contains only plain containers and tensors ({schema_version, model, model_config, git_sha, step, val_loss}) and loads under PyTorch's restricted unpickler — zero code execution on load. Prefer the project's single load path, personacore.checkpoint.load_slim(), which enforces this and validates the schema.

Quickstart

1. Download model_slim.pt from this release into checkpoints/ at the repo root
2. Follow the README quickstart: python scripts/demo_app.py → http://127.0.0.1:7860 (fully offline)

The committed test suite verifies the artifact end-to-end: pytest tests/test_slim_checkpoint.py.