| --- |
| license: cc-by-nc-4.0 |
| tags: |
| - arkadiko |
| - checkpoint |
| - resume-only |
| --- |
| |
| # Arkadiko V4 — Raw Training Checkpoints |
|
|
| **For resume-training and reproducibility only. Not the artifact you want to load for inference or SFT.** |
|
|
| For a clean inference/SFT artifact see [VectorNomad/arkadiko-v4-base](https://huggingface.co/VectorNomad/arkadiko-v4-base). |
|
|
| ## Files |
|
|
| | File | Size | Step | What | |
| |---|---|---|---| |
| | `final.pt` | ~1.28 GB | 9,114,584 | End-of-run checkpoint. Last optimizer step before training terminated. | |
| | `latest.pt` | ~1.28 GB | ~9,113,000 | Atomic-save predecessor. Kept as a safety copy. | |
|
|
| Each `.pt` is a Python pickle (`torch.save`) with these top-level keys: |
|
|
| - `step` — global step |
| - `total_tokens` — tokens consumed |
| - `subphase_idx` — curriculum subphase (15 = final) |
| - `config` — `ArkadikoConfig` dataclass instance (requires the project repo to unpickle) |
| - `model` — state_dict (131 entries, 213.9M bf16 params) |
| - `optimizer` — AdamW state (exp_avg, exp_avg_sq for every param + LR/beta groups) |
|
|
| ## Loading |
|
|
| Requires the `arkadiko` package on `sys.path` (the project repo is private). Without it, `torch.load` raises `ModuleNotFoundError: No module named 'arkadiko'` because the pickled `ArkadikoConfig` references it. |
|
|
| ```python |
| import sys; sys.path.insert(0, "/path/to/arkadiko-repo") |
| import torch |
| ckpt = torch.load("final.pt", map_location="cpu", weights_only=False) |
| print(ckpt["step"], ckpt["total_tokens"]) |
| ``` |
|
|
| For inference-only use of the weights, prefer the `safetensors` artifact in the [base repo](https://huggingface.co/VectorNomad/arkadiko-v4-base). |
|
|
| ## License |
|
|
| CC BY-NC 4.0. |
|
|
