AINovice2005/carbon-likelihood-stats

Dataset Summary

AINovice2005/carbon-likelihood-stats is a derived dataset from the Carbon corpus containing model-based likelihood statistics for biological sequence records.

Each row corresponds to a sequence record identified by record_id and describes the portion of the original corpus associated with that record through its start and end token positions. The dataset captures both aggregate likelihood measurements and statistics describing the distribution of token-level log probabilities within each sequence.

These statistics provide a quantitative view of how the underlying sequence model scores each sequence and can be used for corpus analysis, filtering, sampling, quality-control workflows, and identifying sequences with unusual model likelihood profiles.

Schema

Column	Type	Description
record_id	string	Unique identifier for the source sequence record.
string_lengths	int64	Length of the corresponding sequence/string.
start	int64	Starting token position of the record in the processed/tokenized corpus.
end	int64	Ending token position of the record in the processed/tokenized corpus.
mean_log_prob	float32	Mean token log probability assigned by the model across the evaluated tokens.
sum_log_prob	float32	Sum of the token log probabilities across the evaluated sequence.
perplexity	float32	Sequence perplexity derived from the model's token log probabilities. Lower values indicate that the model assigns higher average probability to the sequence.
supervised_position_count	int64	Number of token positions included in the supervised likelihood calculation.
min_token_logprob	float32	Lowest token-level log probability observed for the sequence.
argmin_position	int64	Token position at which the minimum token log probability occurs.
per_token_logprob_std	float32	Standard deviation of token-level log probabilities within the sequence, describing variation in model confidence across positions.

What the Dataset Represents

The dataset provides sequence-level summaries of model likelihood.

The fields capture several complementary properties:

• mean_log_prob measures the average log probability assigned to tokens.
• sum_log_prob measures the aggregate log probability over the evaluated positions.
• perplexity expresses the average negative log likelihood on the exponential scale.
• min_token_logprob identifies the least probable token position.
• argmin_position locates that lowest-probability token.
• per_token_logprob_std measures how variable the model's confidence is across the sequence.
• supervised_position_count records how many positions contributed to the likelihood statistics.

Together, these fields allow users to examine both the overall likelihood of a sequence and the local variation in model confidence.

Sequence Positioning

start and end provide positional references into the processed/tokenized corpus. They allow the likelihood statistics to be associated with the corresponding region of the underlying corpus.

The values should therefore be treated as corpus/tokenization coordinates, rather than biological coordinates such as genomic chromosome positions.

Intended Uses

This dataset can be used for:

• Model-based sequence quality analysis
• Likelihood distribution analysis
• Sequence ranking and filtering
• Outlier detection
• Corpus sampling
• Identifying sequences containing unusually low-probability regions
• Comparing likelihood distributions across subsets of the Carbon corpus
• Supporting downstream dataset curation

Dataset Statistics

Observed value ranges in the current dataset include:

• string_lengths: 11–14
• mean_log_prob: approximately -9 to -0.05
• sum_log_prob: approximately -131,562 to -17
• perplexity: approximately 1.06–8.1k
• supervised_position_count: 2–16.7k
• min_token_logprob: approximately -30.25 to -6.38