rmems's picture
Cross-link ternary quant experiments + GOZ1 packs (Grok Build: Grok 4.5 high)
f39dbad verified
|
Raw
History Blame Contribute Delete
8.66 kB
---
license: apache-2.0
pretty_name: "xai-dissect Grok-1 structural parse artifacts"
size_categories:
- n<1K
annotations_creators:
- machine-generated
tags:
# Discovery-first (custom tags; avoid format noise like json/jsonl)
- moe
- structural-analysis
- grok-1
- model-dissection
- research-experiment
- personal-research
- mixture-of-experts
- expert-routing
- weight-analysis
- structural-inventory
- checkpoint-inventory
- checkpoint-parse
- model-introspection
- open-weights
- xai
- xai-org/grok-1
- xai-dissect
configs:
- config_name: tensors
data_files:
- split: train
path: data/tensors.jsonl
- config_name: conversion
data_files:
- split: train
path: data/conversion.jsonl
- config_name: routing_critical
data_files:
- split: train
path: data/routing_critical.jsonl
dataset_info:
- config_name: tensors
features:
- name: tensor_index
dtype: int64
- name: shard_path
dtype: string
- name: shard_ordinal
dtype: int64
- name: role
dtype: string
- name: dtype
dtype: string
- name: shape
sequence: int64
- name: nbytes
dtype: int64
- name: kind
dtype: string
- name: block_index
dtype: int64
- name: block_slot
dtype: int64
- name: model_family
dtype: string
- name: checkpoint_ref
dtype: string
splits:
- name: train
num_examples: 770
- config_name: conversion
features:
- name: tensor_index
dtype: int64
- name: tensor_name
dtype: string
- name: structural_name
dtype: string
- name: kind
dtype: string
- name: region
dtype: string
- name: dtype
dtype: string
- name: shape
sequence: int64
- name: numel
dtype: int64
- name: byte_len
dtype: int64
- name: shard_index
dtype: int64
- name: source_shard_path
dtype: string
- name: packing_policy
dtype: string
- name: protected_reason
dtype: string
- name: model_family
dtype: string
- name: checkpoint_ref
dtype: string
splits:
- name: train
num_examples: 770
- config_name: routing_critical
splits:
- name: train
num_examples: 64
---
# Grok-1 structural parse artifacts (personal research)
**Personal research tooling output** — not a product release and not a hybrid-quantization
implementation. Structural parse of open Grok-1 checkpoint shards (tensor inventory, MoE
expert layout, routing-critical tensors, conversion policies, reports) produced by the
[rmems/xai-dissect](https://github.com/rmems/xai-dissect) **weight parser**.
This is **not** a copy of the model weights, **not** a finetune, **not** an inference
runtime, and **does not implement hybrid quantization**.
## Related research (ternary packing / route preservation)
Follow-on personal research from [rmems/grok-ozempic](https://github.com/rmems/grok-ozempic),
packaged by **Grok Build: Grok 4.5 (high)**:
| Dataset | What |
|---------|------|
| [`rmems/grok-1-ternary-quant-experiments`](https://huggingface.co/datasets/rmems/grok-1-ternary-quant-experiments) | Experiment reports, τ sweeps, block-0 forward & multi-block chain metrics |
| [`rmems/grok-1-goz1-packs`](https://huggingface.co/datasets/rmems/grok-1-goz1-packs) | Binary GOZ1 packs (v1 + v3) derived from open Grok-1 |
Collection: [Grok-1 dissection and ternary packing](https://huggingface.co/collections/rmems/grok-1-dissection-and-ternary-packing-6a7927095f6a7ca444f003af).
Those repos measure **route preservation under ternary packing**; this dataset remains
**parser / inventory only** and does not implement hybrid quantization.
## Purpose / research context
These files are **personal experiment artifacts**. They were collected with `xai-dissect` to:
1. Map open [`xai-org/grok-1`](https://huggingface.co/xai-org/grok-1) `ckpt-0` shard layout
(770 tensors, MoE experts, routers, norms).
2. Inventory **packing already present in the open checkpoint** (many expert/attention
shards ship as int8 packs under names like `quant.weight` in the official release).
3. Support **offline personal research** toward hybrid quantization / routing-preserving
packing ideas — as **parser ground truth**, not as a shipped quantizer.
If you are looking for a hybrid quantization system, this dataset is the wrong place: it
is only the dissection metadata from running a personal parser over the public weights.
### About packing labels in the data
| Where | What you will see |
|--------|-------------------|
| **Dataset Viewer** (`data/*.jsonl`) | Viewer-friendly names: `packed_weight`, `packing_policy`, `attention_projection_i8` |
| **Full `artifacts/` tree** | **Native xai-dissect / checkpoint vocabulary** may still say `quant_weight`, `quant_policy`, `quant.weight`, etc. |
Those native strings describe **how the open Grok-1 checkpoint is already packed**, not a
claim that this project applied hybrid quantization.
## Source weights
| Field | Value |
|--------|--------|
| Hub model | [`xai-org/grok-1`](https://huggingface.co/xai-org/grok-1) |
| Checkpoint | `ckpt-0` (770 shards) |
| Portable ref in files | `xai-org/grok-1:ckpt-0` |
```bash
hf download xai-org/grok-1 --include "ckpt-0/*" --local-dir checkpoints
```
Upstream open-weights code: [xai-org/grok-1](https://github.com/xai-org/grok-1).
## Dataset configs (Hub viewer)
| Config | Rows | Description |
|--------|------|-------------|
| `tensors` | 770 | Flat tensor inventory rows (`data/tensors.jsonl`; roles use `packed_weight` for int8 packs) |
| `conversion` | 770 | Per-tensor structural names + `packing_policy` |
| `routing_critical` | 64 | Routing-critical tensor summaries |
```python
from datasets import load_dataset
tensors = load_dataset("rmems/grok-1-dissect-artifacts", "tensors", split="train")
print(tensors[0])
```
## Full artifacts tree
Complete path-scrubbed JSON/Markdown from run `grok1_run3_20260802T023050Z`
(tool-native field names; see packing note above):
```text
artifacts/grok1_run3_20260802T023050Z/
exports/xai-grok-1-ckpt-0/ # inventory, experts, routing, stats (+ findings)
manifests/xai-grok-1-ckpt-0/ # conversion, coverage, routing plans
reports/xai-grok-1-ckpt-0/ # Markdown reports
RUN_ID.txt
```
Also included:
- `provenance.json` — purpose, source model, parser, exclusions, packaging attribution
- `dataset_card.json` — machine-readable card summary
- `file_manifest.json` — path + size + sha256 for every published file
## Architecture snapshot (from inventory)
| Hyperparameter | Value |
|----------------|-------|
| `vocab_size` | 131072 |
| `d_model` | 6144 |
| `n_experts` | 8 |
| `d_ff` | 32768 |
| `n_blocks` | 64 |
| shards / tensors | 770 |
## What is excluded
- Downstream **planning** surfaces that sketched future packing pilots (`quant-plan`,
`pilot-selection-plan`) — those are not part of this personal research drop
- Weight tensors / checkpoint blobs
- Host-absolute paths and usernames (all refs use portable `xai-org/grok-1:ckpt-0/...`)
- Local run logs and comparison notes
- Unrelated local campaign side outputs
## How it was produced
1. Run personal [xai-dissect](https://github.com/rmems/xai-dissect) tooling against open
Grok-1 `ckpt-0` shards.
2. Keep structural parser surfaces only (inventory / experts / routing / stats + manifests).
3. Scrub absolute host paths to portable Hub refs.
4. Emit flat JSONL tables for the Dataset Viewer (viewer-friendly packing field names).
5. Publish as a research dump of parser output — not as a hybrid-quant product.
## Citation
If you use this dataset, please also cite the source weights:
```bibtex
@misc{grok1_dissect_artifacts,
title = {Grok-1 structural parse artifacts (personal research)},
author = {rmems},
year = {2026},
url = {https://huggingface.co/datasets/rmems/grok-1-dissect-artifacts},
note = {Personal xai-dissect parser output; not hybrid quant; weights from xai-org/grok-1}
}
@misc{grok1,
title = {Grok-1},
author = {xAI},
year = {2024},
url = {https://huggingface.co/xai-org/grok-1}
}
```
## License
Apache-2.0 (artifact packaging), matching the open Grok-1 weights license on the Hub.
The parser repository may use a different license; this package contains derived
structural metadata only.
## Attribution
Dataset packaging, path scrubbing, Hub layout, and card authoring:
**Grok Build: Grok 4.5 (high)**.