luca-sartori/degeneration-probe-instruct-token-level

Degeneration Probe Instruct Generations (Token-Level)

Per-token degeneration variant of the instruct generations. chunk_summary is a list of {token_index, degenerating, repetition} records, where repetition is the average bigram-repetition score over all length-256 sliding windows that contain the token, and degenerating is True when that average is >= 0.8.

The rows are sourced from /Users/lucae/Desktop/degeneration/degeneration-probe/outputs/hf_datasets/instruct_concatenated/data (the chunk-level luca-sartori/degeneration-probe-instruct dataset). Every row is preserved verbatim except for two columns: chunk_summary is recomputed at the token level and degenerating is recomputed as "any token in the generation is degenerating". The splits are then re-stratified on (source_dataset, new degenerating label) to maintain balanced positive/negative and source-dataset proportions across train, validation, and test.

Per-token degeneration

For every row, generated_text is retokenized with swiss-ai/Apertus-8B-Instruct-2509 and each token is assigned a repetition score equal to the average bigram (n=2) repetition score over all length-256 sliding windows that contain that token. A token is marked degenerating when this averaged score is >= 0.8. Tokens that lie outside any complete window (only possible when a generation has fewer than 256 tokens) get repetition: null and degenerating: false.

chunk_summary schema (one entry per generated token):

[
  {"token_index": 0, "degenerating": false, "repetition": 0.18181818181818177},
  ...
]

The row-level degenerating flag is True when at least one token has degenerating: True.

Contents

• Rows: 83350
• Tokenizer / model: swiss-ai/Apertus-8B-Instruct-2509
• Sliding window size: 256
• N-gram size: 2
• Token-level degenerating threshold: repetition >= 0.8
• Stratification: source_x_label
• Splits: train, validation, test
• Split fractions: train=0.8, validation=0.1, test=0.1
• Shuffle seed: 42
• Hub repo: luca-sartori/degeneration-probe-instruct-token-level

Split Counts

• train: 66680
• validation: 8335
• test: 8335

Label Counts

• train: positive=915, negative=65765
• validation: positive=113, negative=8222
• test: positive=113, negative=8222

Source Dataset Counts

train

• AI-MO/NuminaMath-1.5: 13360
• FreedomIntelligence/medical-o1-verifiable-problem: 13360
• allenai/IF_sft_data_verified: 13240
• nvidia/Llama-Nemotron-Post-Training-Dataset: 13360
• zwhe99/DeepMath-103K: 13360

validation

• AI-MO/NuminaMath-1.5: 1670
• FreedomIntelligence/medical-o1-verifiable-problem: 1670
• allenai/IF_sft_data_verified: 1655
• nvidia/Llama-Nemotron-Post-Training-Dataset: 1670
• zwhe99/DeepMath-103K: 1670

test

• AI-MO/NuminaMath-1.5: 1670
• FreedomIntelligence/medical-o1-verifiable-problem: 1670
• allenai/IF_sft_data_verified: 1655
• nvidia/Llama-Nemotron-Post-Training-Dataset: 1670
• zwhe99/DeepMath-103K: 1670

Columns

• prompt_id
• prompt
• source_dataset
• model_name
• generated_text
• degenerating (recomputed: any token-level degenerating)
• prompt_mode
• max_new_tokens
• temperature
• top_p
• generation_seed
• prompt_generation_seed
• chunk_summary (per-token list of {token_index, degenerating, repetition})
• dataset_run
• shard
• source_file

Usage

from datasets import load_dataset

ds = load_dataset("luca-sartori/degeneration-probe-instruct-token-level")
print(ds)
print(ds["train"][0]["chunk_summary"][:5])