metadata
dataset_info:
features:
- name: depth
dtype: int64
- name: width
dtype: int64
- name: tokens
dtype: int64
- name: FLOPs_per_token
dtype: float64
- name: FLOPs
dtype: float64
- name: params
dtype: float64
- name: params_with_embeds
dtype: float64
- name: FLOPs_6N
dtype: float64
- name: params_pred_loss
dtype: float64
- name: wd_ratio
dtype: float64
- name: wd_pred_loss
dtype: float64
- name: bucket
dtype: string
splits:
- name: train
num_bytes: 1772
num_examples: 13
download_size: 6825
dataset_size: 1772
configs:
- config_name: default
data_files:
- split: train
path: >-
mins_1e-3/mins_lr_ablation_hot_width_depth_params_relaxed_params/train-*
license: mit
This dataset is my cache for the scaling-laws related to the gemstone models.
In data_cache is the approach 3 data cache with the mins for delta=1e-4, the mins for delta=1e-3 are in mins_1e-3.
This is the code I used to upload it:
import pandas as pd
from datasets import Dataset
import os
import gc
def get_data_dict(path):
contents = os.listdir(path)
ds_store = {}
for i, file in enumerate(contents):
gc.collect()
df = pd.read_parquet(f"{path}{file}")
for col in df.columns:
if pd.api.types.is_interval_dtype(df[col]):
df[col] = df[col].astype(str)
hf_dataset = Dataset.from_pandas(df)
ds_store[file.replace(".parquet", "")] = hf_dataset
hf_dataset.push_to_hub(
"smcleish/scaling-laws-cache",
private=True,
data_dir=path.split("/")[1] + "/" + file.replace(".parquet", ""),
)
gc.collect()
ds_1 = get_data_dict("plotters/data_cache/")
ds_2 = get_data_dict("plotters/mins_1e-3/")
To download it do the oppostite of this. The cache is very large, so maybe target specific files you would like. The approach 3 code is expecting pandas .parquet files.
Please open a discussion with any questions as this is currently very experimental.