File size: 6,353 Bytes
bef160c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 | ---
license: cc-by-4.0
task_categories:
- other
language:
- en
size_categories:
- n<1K
tags:
- weibull
- transformer
- weight-statistics
- model-diagnostics
- npm-weibull
pretty_name: NPM-Weibull DATABASE v9_1
configs:
- config_name: default
data_files: DATABASE_v9_1.csv
---
# NPM-Weibull DATABASE v9_1
Benchmark database of Weibull (k, λ) fits for **12 transformer model entries** spanning **7 architectural families** (Pythia, OLMo-1/2, LLaMA-3, Mistral, Qwen2.5, Qwen3) — 70M–14B parameters, GeLU and SwiGLU activations, Pre-LN and QK-Norm placements.
- 📄 **Paper**: [A Two-Parameter Weibull Framework for Diagnosing Transformer Weight Distributions](https://arxiv.org/abs/2605.18898) (arXiv:2605.18898)
- 📦 **Source repo (primary)**: [github.com/tiexinding/NPM-Weibull-public](https://github.com/tiexinding/NPM-Weibull-public) — directory `database_v9_1/`
- 🔧 **Python library**: `pip install npm-weibull-py` — [PyPI](https://pypi.org/project/npm-weibull-py/)
---
## 📌 Source of truth
The **primary, authoritative location** of this dataset is the GitHub repository above (directory `database_v9_1/`). This Hugging Face Dataset is a **mirror** synced from GitHub on each tagged release, provided for `load_dataset(...)` convenience and Dataset Viewer discoverability.
- Updates are **batched per release**, not per commit — for the latest state, watch the GitHub repo.
- Issues and pull requests: please file on [GitHub Issues](https://github.com/tiexinding/NPM-Weibull-public/issues) rather than HF discussions (faster response).
- License, citation, and documentation are kept consistent across both surfaces.
---
## Quick start
### Option 1 — `load_dataset` (Hugging Face)
```python
from datasets import load_dataset
ds = load_dataset("TiexinDing/NPM-Weibull-DATABASE-v9_1")
print(ds["train"][0]) # First entry: pythia-70m
```
### Option 2 — pandas
```python
import pandas as pd
df = pd.read_csv("DATABASE_v9_1.csv")
print(df[["entry_id", "k_median_o", "lambda_median_o"]].head())
```
### Option 3 — `npm-weibull-py` library (recommended for analysis)
```python
from npm_weibull import DATABASE_v9_1, compare_to_benchmark
print(DATABASE_v9_1) # 12 entries with per-component fits
cmp = compare_to_benchmark({
"median_k_per_kind": {"q": 1.14, "k": 1.13, "v": 1.19, "o": 1.19}
})
print(cmp["nearest_neighbor"]) # nearest of the 12 benchmark entries
```
---
## Dataset structure
### 12 entries
| `entry_id` | Family | Size | Architecture | QK-Norm | Tokens |
|---------------|----------|-------|---------------|---------|--------|
| pythia-70m | Pythia | 70m | MHA-merged | False | 300B |
| pythia-160m | Pythia | 160m | MHA-merged | False | 300B |
| pythia-410m | Pythia | 410m | MHA-merged | False | 300B |
| pythia-1b | Pythia | 1B | MHA-merged | False | 300B |
| pythia-6.9b | Pythia | 6.9B | MHA-merged | False | 300B |
| olmo-1-7b | OLMo-1 | 7B | MHA-separate | False | 2.5T |
| olmo-2-7b | OLMo-2 | 7B | MHA-separate | True | 5T |
| llama-3-8b | LLaMA-3 | 8B | GQA-4:1 | False | 15T |
| mistral-7b | Mistral | 7B | GQA-4:1 | False | 8T |
| qwen2.5-7b | Qwen2.5 | 7B | GQA-7:1 | False | 18T |
| qwen2.5-14b | Qwen2.5 | 14B | GQA-5:1 | False | 18T |
| qwen3-8b | Qwen3 | 8B | GQA-4:1 | True | 36T |
### Columns (55 total)
**Identifiers**: `entry_id`, `family`, `size`, `arch`
**Architecture**: `n_q`, `n_kv` (head counts for GQA), `qk_norm`
**Training hyperparameters**: `training_tokens`, `eta_peak` (peak learning rate), `lambda_wd` (weight decay), `T_steps`, `tau_iter = 1/(η·λ_wd)`, `T_over_tau` (Wang-Aitchison 2024 cycle ratio), `Physical_State`, `hp_confidence` (explicit / inferred / estimated)
**Per-component Weibull fits** (median across blocks within each model):
- Pythia entries (GeLU 2-projection FFN with merged W_qkv): populate `qkv`, `o`, `ffn_in`, `ffn_out`
- Non-Pythia entries (SwiGLU 3-projection FFN with separate Q/K/V): populate `q`, `k`, `v`, `o`, `gate`, `up`, `down`
Each component has four columns: `k_median_*`, `lambda_median_*`, `R2_median_*`, `R2_below_95_*` (count of blocks with R² < 0.95).
`n_records` = total number of per-block Weibull fits aggregated.
---
## Key findings (paper §3)
1. **Transmission Class** (FFN + W_o): the **median terminal k across components per entry**, then aggregated across the 12 entries, falls in **[1.186, 1.204]** with cross-family CV = **0.51%**. Shared across SwiGLU and GeLU activations, Pre-LN and QK-Norm placements, and model sizes 70M → 14B.
2. **Selection Class** (W_q, W_k): k departs from initialization anchor (k_init ≈ 1.205), severity tracks attention storage architecture — **separately-stored MHA** (OLMo-1/2) deepest at k ∈ [0.76, 0.99], **GQA** (LLaMA-3, Mistral, Qwen2.5/3) milder at [1.10, 1.16], **Pythia merged W_qkv** transitional at [1.05, 1.18].
3. **λ scaling**: terminal λ ∝ √(η/λ_wd) within Pythia, Pearson r = 0.94 across the three Transmission Class component kinds, directionally consistent with Fan et al. (2025) AdamW steady-state analysis.
---
## Files in this dataset
| File | Purpose |
|---|---|
| `DATABASE_v9_1.csv` | Machine-readable dataset (55 columns × 12 entries) |
| `DATABASE_v9_1.md` | Human-readable reference table |
| `DATABASE_v9_1_report.md` | Per-entry sanity verification |
| `populate_database_v9_1.py` | Internal regeneration script (dev-only; requires cascade v3 raw per-block fits not shipped here — see GitHub repo for the cascade pipeline) |
| `LICENSE` | CC BY 4.0 license text |
| `README.md` | This file |
---
## License
[CC BY 4.0](LICENSE) — free for academic and commercial use with attribution.
---
## Citation
```bibtex
@article{ding2026twoparameterweibull,
title = {A Two-Parameter Weibull Framework for Diagnosing Transformer Weight Distributions},
author = {Ding, Tiexin},
journal = {arXiv preprint arXiv:2605.18898},
year = {2026},
doi = {10.48550/arXiv.2605.18898},
url = {https://arxiv.org/abs/2605.18898}
}
```
---
**Author**: Tiexin Ding · NeuralCAE
Issues / questions: [GitHub Issues](https://github.com/tiexinding/NPM-Weibull-public/issues)
|