fin-abandoned

This is the final saved VEG hybrid checkpoint from the abandoned Wind Flame 2.0 run. It is an unfinished research artifact, not a promoted or finished model. Training was stopped at the owner’s request on November 6, 2026, after 1,796 of 3,000 planned recovery updates. No agent benchmark or untouched test-split evaluation was completed for this hybrid; no 4B-class performance claim has been established.

The model has approximately 383.2 million parameters: one transformer prelude layer, 14 VEG core layers executed once, and one transformer coda layer. It was initialized from a 246M VEG parent and trained using teacher KL plus cross-entropy on text data. At the final saved update the recorded validation losses were 3.3034 on capability and 3.7643 on nitron. These are text losses, not agent benchmark scores. The included training log records the run.

Files and loading

• checkpoint.pt is the original final PyTorch checkpoint, SHA-256 15fee504506c6b6ec68116feaa8dbe583eb9205b91b2e61cd9b2b0ab88355976.
• veg/ contains the corresponding VEG implementation; tokenizer.json is its 32k BPE tokenizer.
• Dependencies include PyTorch, NumPy and tokenizers. This is custom research code, not a Transformers AutoModel checkpoint. PyTorch pickle checkpoints must only be loaded from trusted sources.

The checkpoint was trained with steps=1. Its saved configuration defaults to four steps, so always pass steps=1 explicitly. Four passes produce a different, unvalidated model. For example, from this repository directory:

import torch
from veg.config import VAREConfig
from veg.model import VAREModel
from veg.tokenizer import BPETokenizer

ck = torch.load("checkpoint.pt", map_location="cpu", weights_only=False)
model = VAREModel(VAREConfig.from_dict(ck["config"]))
model.load_state_dict(ck["model"], strict=True)
model.eval()
tokenizer = BPETokenizer("tokenizer.json")
ids = torch.tensor([tokenizer.encode("Hello")])
with torch.inference_mode():
    logits = model(ids, steps=1).logits

This archive is provided as an abandoned experimental checkpoint; behavior, safety, and downstream suitability have not been established.