prepretraining-training-samples-v1

First 10 training sequences (4096 tokens each) for each of the 4 conditions, decoded back to human-readable text. Shows exactly what the model sees during training. Use this to verify data quality, ordering, and condition differentiation.

Dataset Info

• Rows: 40
• Columns: 8

Columns

Column	Type	Description
condition	Value('string')	Data scheduling condition: baseline, front-load, constant-mix, or anneal
condition_description	Value('string')	What this condition feeds the model during its first phase
sequence_index	Value('int64')	Index of this sequence within the condition (0 = very first thing the model sees)
text	Value('string')	Decoded text of the 4096-token training sequence (human-readable)
n_tokens	Value('int64')	Number of tokens in this sequence (always 4096)
n_chars	Value('int64')	Character count of the decoded text
n_documents	Value('int64')	Number of documents packed into this sequence (separated by EOS tokens)
source_files	Value('string')	Source .npy file names (first 3)

Generation Parameters

{
  "script_name": "analysis/sample_training_batches.py",
  "model": "N/A (data inspection, not model output)",
  "description": "First 10 training sequences (4096 tokens each) for each of the 4 conditions, decoded back to human-readable text. Shows exactly what the model sees during training. Use this to verify data quality, ordering, and condition differentiation.",
  "hyperparameters": {
    "sequence_length": 4096,
    "n_sequences_per_condition": 10,
    "tokenizer": "allenai/gpt-neox-olmo-dolma-v1_5"
  },
  "input_datasets": [
    "reasoning-degeneration-dev/prepretraining-gold-v1",
    "reasoning-degeneration-dev/prepretraining-web-v1"
  ],
  "experiment_id": "prepretraining",
  "artifact_type": "input_data",
  "visualizer_type": "table",
  "artifact_group": "data-inspection"
}

Experiment Documentation

For complete experiment details, see https://github.com/Zayne-sprague/SC-Research-Notes/tree/main/experiments/prepretraining

Usage

from datasets import load_dataset

dataset = load_dataset("reasoning-degeneration-dev/prepretraining-training-samples-v1", split="train")
print(f"Loaded {len(dataset)} rows")

---

This dataset is tracked in reasoning-degeneration-dev/PROJECT-MANIFEST