---
license: cc-by-4.0
pretty_name: "4D LLM Scaling-Law Grid Run Metrics"
tags:
- scaling-laws
- llm
- language-model-pretraining
- training-dynamics
- learning-curves
size_categories:
- 100K<n<1M
---

# 4D LLM Scaling-Law Grid: Run Metrics

Per-step training metrics from a four-dimensional grid of small language-model pretraining runs, produced for the scaling-law study "Tokens-per-Parameter Coverage Is Critical for Robust LLM Scaling Law Extrapolation". The archive was moved here from the [`lab-v2/4d_llm_grid`](https://github.com/lab-v2/4d_llm_grid) code repository to keep the large binary metric files out of Git.

## Contents

`metrics/` holds **462 gzipped CSV files** (~2.4 GB), one per training run. Filenames encode each run's position on the grid:

```
grid4d_row{idx}_N{n_params}_D{token_budget}_L{n_layers}d{d_model}h{n_heads}_lr{lr}_wd{wd}.csv.gz
```

## The four axes

Every run varies along four dimensions:

- **N** - model size / shape (`n_layers`, `d_model`, `n_heads`)
- **D** - token budget (training-data size)
- **lr** - peak learning rate
- **wd** - weight decay

Models are decoder-only LLaMA-style transformers (`max_seq_len = 256`) pretrained on a subset of C4.

## File format

Each decompressed `.csv.gz` has:

1. A leading comment line `# config: {...}`, a JSON object with the full run configuration: `n_params`, `n_layers`, `d_model`, `n_heads`, `max_seq_len`, `tpp`, `tpp_eff`, `token_budget`, `lr`, `weight_decay`, `per_gpu_batch_size`, `world_size`, `est_total_steps`, `dataset`.
2. A CSV table with columns:

| column | description |
| --- | --- |
| `step` | optimizer step |
| `tokens` | tokens seen so far |
| `train_loss` | training loss |
| `val_loss` | validation loss (where logged) |
| `grad_norm` | gradient norm |
| `lr` | learning rate at this step |

Runs are roughly 2,000 steps each; the collection totals on the order of 10^6 per-step rows.

## Usage

```python
import gzip, json
import pandas as pd
from huggingface_hub import hf_hub_download

fn = "metrics/grid4d_row0000_N5000000_D100000000_L2d48h1_lr0.00032_wd0.05.csv.gz"
path = hf_hub_download("leibnitz-lab/4d-llm-grid-metrics", fn, repo_type="dataset")

with gzip.open(path, "rt") as f:
    config = json.loads(f.readline().removeprefix("# config: "))  # first line is a JSON config comment
    df = pd.read_csv(f)

print(config["tpp"], config["lr"], config["weight_decay"])
print(df.columns.tolist(), len(df))
```

## Provenance

Generated by the code at [`lab-v2/4d_llm_grid`](https://github.com/lab-v2/4d_llm_grid); see that repository's `STATUS.md` and `analysis/` for the grid definition, coverage, and downstream analysis.

## License

Released under [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/).

## Citation

```bibtex
@misc{tpp_coverage_4d_grid_metrics,
  title = {Tokens-per-Parameter Coverage Is Critical for Robust LLM Scaling Law Extrapolation},
  note  = {Dataset: 4D LLM scaling-law grid run metrics},
  year  = {2026}
}
```