Datasetpublic

rmems/grok-1-goz1-packs

Grok-1 GOZ1 ternary weight packs (research) Packaged by: Grok Build: Grok 4.5 (high) Research binary packs produced by rmems/grok-ozempic from open xai-org/grok-1 ckpt-0 weights (Apache-2.0). Format is GOZ1 (custom ternary/SNN container) — not safetensors, not GGUF, not loadable by transformers. Companion experiment reports & metrics: rmems/grok-1-ternary-quant-experiments. Structural inventory (no weights): rmems/grok-1-dissect-artifacts. What is included… See the full description on the dataset page: https://huggingface.co/datasets/rmems/grok-1-goz1-packs.

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes175downloads

Dataset Card

Grok-1 GOZ1 ternary weight packs (research)

Packaged by: Grok Build: Grok 4.5 (high)

Research binary packs produced by rmems/grok-ozempic from open `xai-org/grok-1` ckpt-0 weights (Apache-2.0). Format is GOZ1 (custom ternary/SNN container) — not safetensors, not GGUF, not loadable by transformers.

Companion experiment reports & metrics: `rmems/grok-1-ternary-quant-experiments`.

Structural inventory (no weights): `rmems/grok-1-dissect-artifacts`.

What is included

Campaign	Contents	GOZ1 versions on disk
`packs/grok1-first-embed.goz1`	First embedding quant (#39)	v1
`packs/goz66-embed-v3.goz1`	Embedding pack with GOZ1 v3 metadata	v3
`packs/goz66-block000-tier-v3.goz1`	Block-0 tiered policy pack	v3
`packs/tau-sweep/`	Embedding τ sweep packs + trit histograms	v1
`packs/block-pilot/`	attention-only / expert-only / full block packs + route metrics	v1
`packs/block0-forward-tau/`	Attention τ packs for genuine block-0 forward	v1
`packs/multiblock-68/`	Blocks 0–3 attention+expert packs for multi-block chain	v3

Format versions

Version	Fields beyond v1
v1	magic, tensor table, trit/f16 payloads
v2	per-tensor `scale` + sentinel `0x5CA1E021` (reader still supports; no standalone v2 files retained on this host)
v3	v2 + `gif_threshold` + `threshold_abs` per tensor row

Integrity

See pack_inventory.json for per-file sha256, byte size, and GOZ1 version.

The Dataset Viewer indexes 27 inventory rows from data/packs.jsonl (packs / train). That is the example count (size_categories: n<1K). The binary GOZ1 payload is ~11 GB logical (total_bytes 11002996704); Hub LFS usedStorage is lower because some packs share SHA-256.

python

from datasets import load_dataset
rows = load_dataset("rmems/grok-1-goz1-packs", "packs", split="train")
print(rows[0])

Download examples

bash

hf download rmems/grok-1-goz1-packs --include "pack_inventory.json" "data/*" "README.md"
hf download rmems/grok-1-goz1-packs --include "packs/goz66-embed-v3.goz1" "packs/grok1-first-embed.goz1"
hf download rmems/grok-1-goz1-packs --include "packs/multiblock-68/*"

License & attribution

—Packs are derived from xai-org/grok-1 (Apache-2.0).
—Packaged by Grok Build: Grok 4.5 (high).

Citation

bibtex

@misc{grok1_goz1_packs,
  title  = {Grok-1 GOZ1 ternary weight packs (research)},
  author = {rmems},
  year   = {2026},
  note   = {Packaged by Grok Build: Grok 4.5 (high)},
  url    = {https://huggingface.co/datasets/rmems/grok-1-goz1-packs}
}