Skip to main content

Preparing Datasets

Format your data correctly before training with ColabHive. This guide covers the supported formats, examples, and the data-volume guidance you should size against.

Supported formats​

The direct-upload dataset formats are csv, jsonl, and parquet. When you upload a file, the format is detected from its extension:

ExtensionStored as
.csvcsv
.jsonl or .jsonjsonl
.parquetparquet

Use client.datasets.upload() for the supported single-file upload path. The API accepts files up to 10 GiB, subject to the gateway and scheduler memory available in the target installation.

Tabular data (regression & classification)​

Use CSV, JSONL, or Parquet for XGBoost, LightGBM, Random Forest, CatBoost, SVM, MLP, TabNet, and BERT classification.

Simple, human-readable. Column headers on the first row.

age,income,loan_amount,credit_score,approved
35,55000,10000,720,1
42,80000,25000,680,1
28,30000,5000,580,0

Rules:

  • First row must be column headers.
  • Feature columns: any numeric or text values.
  • Target column: the column to predict (specify in hyperparameters).
  • Missing values: leave blank — ColabHive handles them automatically.
  • Minimum row counts vary by model — see Data volume guidance.

JSONL (JSON Lines)​

One JSON object per line. Good for mixed types and nested structures.

{"age": 35, "income": 55000, "loan_amount": 10000, "credit_score": 720, "approved": 1}
{"age": 42, "income": 80000, "loan_amount": 25000, "credit_score": 680, "approved": 1}
{"age": 28, "income": 30000, "loan_amount": 5000, "credit_score": 580, "approved": 0}

Parquet​

Columnar format. Best for large datasets (millions of rows) thanks to compression and fast reads.

import pandas as pd

df = pd.read_csv("data.csv")
df.to_parquet("data.parquet", index=False)

Specifying the target column​

By default, ColabHive uses the last column as the target. Override it:

job = client.training.create(
    model="xgboost-regression",
    dataset_id=dataset.id,
    hyperparameters={
        "target_column": "approved",  # specify explicitly
        "n_estimators": 100,
    },
)

Multi-output regression​

XGBoost, LightGBM, Random Forest, MLP, and TabNet support predicting multiple target columns in a single run:

hyperparameters = {
    "target_columns": ["temp", "humidity", "pressure"],  # multiple targets
}

LLM fine-tuning datasets​

For llm-qlora-finetune (transformers + PEFT backend). The format depends on the dataset_type hyperparameter. Full parameter reference: LLM Fine-Tuning.

Alpaca (dataset_type: "alpaca") — recommended​

Best for instruction-following and Q&A. Loss is computed only on output.

{"instruction": "Write a Python function to reverse a string", "input": "", "output": "def reverse_string(s):\n    return s[::-1]"}
{"instruction": "What is the capital of France?", "input": "", "output": "Paris"}
  • instruction (required), input (optional context), output (required).

Chat (dataset_type: "chat") — multi-turn / ChatML​

messages ([{role, content}]) or sharegpt-style conversations ([{from, value}]), rendered with the tokenizer's chat template.

{"messages": [{"role": "user", "content": "Hello, how are you?"}, {"role": "assistant", "content": "Doing great, thanks!"}]}
{"messages": [{"role": "system", "content": "You are a coding assistant."}, {"role": "user", "content": "What is a closure?"}, {"role": "assistant", "content": "A closure captures variables from its enclosing scope..."}]}

Text (dataset_type: "text") — raw completion​

A single free-text column (defaults to text; override with text_column).

{"text": "The quick brown fox jumps over the lazy dog. This is an example of a pangram."}
{"text": "Machine learning is a subset of AI that enables computers to learn from data."}

Time series datasets​

For timesfm-2.5-finetune-gpu, prophet-forecasting, arima-forecasting, classical-forecasting-auto, and the other forecasters.

A CSV with a date/timestamp column and one or more numeric target columns:

date,sales,temperature
2023-01-01,1200,22.5
2023-01-02,1350,23.1
2023-01-03,1100,21.8

Requirements:

  • Date column consistently formatted (YYYY-MM-DD recommended).
  • Regular intervals (daily, weekly, monthly — no gaps).
  • Target column(s) numeric.
  • Length: at least 24 points for ARIMA; more for Prophet and TimesFM (see Data volume guidance).
hyperparameters = {
    "date_column": "date",
    "target_column": "sales",                      # single target
    # or:
    "target_columns": ["sales", "temperature"],    # multi-target
}

Text classification datasets (BERT)​

For bert-classification-gpu:

text,label
"This movie was absolutely fantastic!",positive
"Terrible service, never going back.",negative
"Average experience, nothing special.",neutral
  • text: the text to classify (any length).
  • label: the class label (string or integer).
  • At least 100 samples per class (see Data volume guidance).

Best practices​

  • No duplicate rows. Remove exact duplicates before uploading.
  • Consistent missing values. Empty strings for CSV, null for JSONL — don't mix "N/A", 0, -999.
  • Balanced classes (classification). Extreme imbalance (>10:1) hurts performance.
  • No target leakage. Don't include columns derived from or trivially correlated with the target.

Data volume guidance​

These are the single reference minimums for this guide. They are guidance for good results, not hard limits enforced by the API.

Model familyMinimumSweet spot
Linear / Logistic Regression100 rows500 – 10,000
Random Forest200 rows1,000 – 100,000
XGBoost / LightGBM500 rows10,000 – 1,000,000
MLP / TabNet1,000 rows100,000+
BERT classification100 / class1,000 – 50,000
LLM fine-tuning (alpaca)500 rows2,000 – 10,000
LLM fine-tuning (chat)200 rows1,000 – 5,000
LLM domain fine-tuning1,000 rows5,000 – 50,000
Time series (ARIMA)24 points100 – 10,000
Time series (Prophet)50 points100 – 10,000
Time series (TimesFM)100 points500 – 10,000

See also​