README.md

metadata

license: apache-2.0
library_name: pytorch
tags:
  - dynamic-chunking
  - tokenization
  - hnet
  - byte-level
  - beetle-hnet

ali_pkm_s0

A byte-level H-Net run from beetle-hnet (architecture=pkm, objective=ratio schedule=None, dechunk=ema, seed 0, 6000 steps x 32 x 1024 bytes, langs ['en', 'tr', 'fi']).

Layout (Beetle-style)

• step-N branches, one per kept checkpoint on a phase-aware log2 schedule (0, 1, 2, 4, 6, 8, 16, ..., 2^k from every objective-phase start, plus the final step). Each holds model.safetensors, config.json, training_config.yaml and the full resumable trainer state training_state/checkpoint.pt (for inventory arms this includes the inventory).
• main: the latest step's weights and config, plus log.jsonl, final.json and the FLORES+ boundary probes probe/ once the run has finished.

from huggingface_hub import hf_hub_download
import torch
p = hf_hub_download("AdaptiveChunking/ali_pkm_s0", "training_state/checkpoint.pt", revision="step-1024")
state = torch.load(p, map_location="cpu", weights_only=False)

model.safetensors stores each tied tensor once; config.json["tied_weights"] maps the dropped names to the stored ones.