AINovice2005/carbon-tokenized-corpus

Dataset Summary

AINovice2005/carbon-tokenized-corpus is the tokenized sample of AINovice2005/carbon-cpu-enriched-sequences-sampled .

Schema

The current dataset contains the following fields:

Field	Type	Description
record_id	string	Source/reference sequence identifier
start	int64	Start coordinate of the sequence interval
end	int64	End coordinate of the sequence interval
token_ids	list	Integer token IDs produced by the tokenizer
token_mask	list	Per-token mask aligned with token_ids
token_length	int32	Number of tokens associated with the sequence

The Hub currently reports a single train split with approximately 3.07M rows.

Sequence Identity and Lineage

The tuple:

(record_id, start, end)

is retained as the primary sequence-interval identity.

This allows tokenized records to be aligned with their corresponding sequence intervals in upstream and downstream Carbon datasets.

The tokenized dataset therefore preserves the genomic coordinate lineage even though the original biological feature columns are not duplicated into every tokenized row.

For cross-stage analysis, biological metadata should be joined from the appropriate upstream dataset using the sequence identity rather than assumed to be present in this artifact.

Token Representation

token_ids contains the integer representation generated by the Carbon tokenization pipeline.

The exact interpretation of individual token IDs depends on the tokenizer/model configuration used to produce the dataset. Token IDs should therefore not be interpreted as nucleotide values or biological annotations directly.

token_mask is stored alongside token_ids and is aligned element-by-element with the token sequence.

The dataset also materializes token_length, allowing token-count statistics to be computed without repeatedly materializing the token-ID arrays.

Intended Uses

This dataset is intended for:

1. Model input preparation

   • efficient downstream model processing
   • likelihood estimation
   • embedding generation

2. Tokenization analysis

   • sequence-length vs token-length relationships
   • token-count distributions
   • tokenization coverage
   • identification of unusually short or long tokenized sequences

3. Cross-stage analysis

   • joining token counts to CPU-derived biological properties
   • relating token length to likelihood statistics
   • evaluating tokenization behavior across biological populations

4. Compute planning

   • estimating downstream model-processing cost
   • calculating token budgets
   • characterizing the computational population before GPU processing

Analytical Role

This dataset is primarily a model-input and tokenization layer, rather than the main biological analytical corpus.

Its principal analytical questions are:

How does the biological sequence population translate into model tokens?

and:

How do token counts vary with sequence properties and biological populations?

For example, token-length distributions can be compared against sequence length, GC content, coding status, taxonomy, and other CPU-derived features.

Coverage

The current lineage shows approximately:

CPU-enriched sequences       3,221,596
             │
             ▼
Tokenized sequences          3,070,722

This corresponds to approximately 95.3% retention between the CPU-enriched and tokenized stages in the current lineage analysis.

The remaining sequences should not automatically be interpreted as biologically invalid; coverage differences can arise from tokenization-stage validity and processing constraints.

Relationship to Likelihood and Embeddings

The tokenized corpus is an upstream computational artifact for model-derived enrichment:

                     tokenized sequence
                            │
             ┌──────────────┴──────────────┐
             ▼                             ▼
      model likelihood                 embeddings
             │                             │
             ▼                             ▼
 carbon-likelihood-stats          carbon-embeddings

The likelihood and embedding datasets represent downstream measurements over the population that actually reaches GPU/model processing.

Consequently, the number of rows in those datasets should not be interpreted as a simple measure of tokenization quality.

Macro vs Matched Analysis

Two forms of analysis should be distinguished.

Macro tokenization statistics

Use the full valid tokenized population to characterize:

• token-length distribution
• token-length vs sequence-length relationship
• total token volume
• tokenization coverage
• token-count extremes

These describe the tokenized dataset itself.

Matched cross-modal analysis

When token counts are compared directly with biological features or model-derived values, use the matching sequence identity:

(record_id, start, end)

This allows tokenization, CPU-derived biological properties, likelihood statistics, and embeddings to be compared on the same sequence intervals when those measurements are all available.

Example

row = ds["train"][0]

print(row["record_id"])
print(row["start"], row["end"])
print(row["token_length"])
print(row["token_ids"][:20])
print(row["token_mask"][:20])

Recommended Analytical Workflow

For corpus-level analysis:

carbon-pilot-corpus-sampled
        │
        ├── biological macro statistics
        │
        ▼
carbon-tokenized-corpus
        │
        ├── tokenization macro statistics
        │
        ▼
likelihood / embedding datasets
        │
        └── model-derived analysis

For cross-modal analysis, retain the common sequence identity:

(record_id, start, end)

and join only the measurements required for the specific analysis.

This avoids treating datasets with different coverage as though they represented identical populations.