luca-sartori/degeneration-probe-base

Degeneration Probe Base Generations

This dataset contains base-model generations produced for the degeneration probe experiments. All generation shards from outputs/generations/base were concatenated, shuffled reproducibly, split into train, validation, and test, and stored as Parquet.

Contents

• Rows: 60529
• Source files: 444 generations.jsonl shards
• Model: swiss-ai/Apertus-8B-2509
• Splits: train, validation, test
• Split fractions: train=0.8, validation=0.1, test=0.1
• Shuffle seed: 42
• Stratification: source_dataset x degenerating
• Hub repo: luca-sartori/degeneration-probe-base

Split Counts

• train: 48423
• validation: 6053
• test: 6053

Label Counts

• train: positive=7884, negative=40539
• validation: positive=985, negative=5068
• test: positive=986, negative=5067

Source Dataset Counts

train

• AI-MO/NuminaMath-1.5: 10847
• FreedomIntelligence/medical-o1-verifiable-problem: 13413
• allenai/IF_sft_data_verified: 2005
• nvidia/Llama-Nemotron-Post-Training-Dataset: 12901
• zwhe99/DeepMath-103K: 9257

validation

• AI-MO/NuminaMath-1.5: 1355
• FreedomIntelligence/medical-o1-verifiable-problem: 1677
• allenai/IF_sft_data_verified: 251
• nvidia/Llama-Nemotron-Post-Training-Dataset: 1612
• zwhe99/DeepMath-103K: 1158

test

• AI-MO/NuminaMath-1.5: 1356
• FreedomIntelligence/medical-o1-verifiable-problem: 1676
• allenai/IF_sft_data_verified: 251
• nvidia/Llama-Nemotron-Post-Training-Dataset: 1614
• zwhe99/DeepMath-103K: 1156

Columns

• prompt_id
• prompt
• source_dataset
• model_name
• generated_text
• degenerating
• prompt_mode
• max_new_tokens
• temperature
• top_p
• generation_seed
• prompt_generation_seed
• chunk_summary
• dataset_run
• shard
• source_file

Usage

from datasets import load_dataset

ds = load_dataset("luca-sartori/degeneration-probe-base")
print(ds)
print(ds["train"][0])

Notes

The text content is not deduplicated or rewritten. The added degenerating, dataset_run, shard, and source_file columns preserve labels and original shard provenance. The shuffle, stratified split, and within-split positive/negative interleaving are deterministic for the configured seed.