Carbon

What is Carbon?

Carbon is a family of autoregressive genomic foundation models from Hugging Face Bio for DNA generation, sequence likelihood scoring, and reference-versus-alternate sequence comparison. It treats DNA as a language modeling problem, but uses a DNA-specific 6-mer tokenizer instead of ordinary text tokenization for bases.

ProteinIQ runs Carbon on canonical DNA sequences containing only A, C, G, and T. Ambiguity bases such as N, degenerate IUPAC codes, RNA U, and non-sequence annotations are not valid Carbon input.

Carbon is most useful when the question depends on learned genomic sequence context: extending a DNA sequence, ranking sequences by model likelihood, or asking whether an alternate sequence is more or less likely than a reference under the selected model. For deterministic sequence editing, use DNA mutator, DNA shuffle, or Reverse complement instead.

How to use Carbon online

Run Carbon online by pasting canonical DNA or uploading FASTA or text files, then choosing generation, scoring, or compare mode. ProteinIQ runs the selected Hugging Face Carbon checkpoint and returns generated DNA, likelihood scores, reference-versus-alternate deltas, spreadsheet data, and downloadable CSV or JSON files. FNS scoring also returns the full per-base probability distribution.

Inputs

Input	Accepted values	Notes
DNA sequence	Raw DNA, FASTA, .fa, .fasta, or .txt	Only uppercase or lowercase canonical DNA bases are accepted after normalization to A, C, G, and T.
Reference sequence	Canonical DNA	Required in Compare mode through the dedicated reference input.
Alternate sequence	Canonical DNA	Required in Compare mode through the dedicated alternate input.

Generation and standard token scoring require at least 6 bp because each DNA token represents one non-overlapping 6-mer. FNS base-pair scoring accepts sequences from 1 bp and preserves partial final 6-mers through Carbon's native right-padding and length recovery. Sequences longer than the selected context window are scored or extended using the rightmost context segment. The maximum accepted sequence length is 786,432 bp.

Modes

Mode	What it does	Typical use
Generation	Extends the submitted DNA sequence by up to Max new bp.	Designing short continuations from a known genomic context or exploring model-preferred next bases.
Score	Calculates model likelihood for each submitted sequence.	Ranking sequence candidates or checking whether edited sequences remain plausible under Carbon.
Compare	Scores a reference and alternate sequence, then reports the alternate-minus-reference delta.	Variant-style analysis, motif perturbation checks, or comparing two candidate edits.

In Compare mode, enter the sequences in the dedicated Reference sequence and Alternate sequence inputs. Their roles are explicit, so Carbon does not infer comparison order from unrelated inputs or settings.

Model choices

Model	Best fit	Context behavior
Carbon-500M	Fast drafts and lower-cost exploratory runs.	Smaller native context than the larger models.
Carbon-3B	Default model for most Carbon jobs.	Flagship checkpoint with a 32,768-token native DNA context, about 196 kbp.
Carbon-8B	Higher-capacity runs when runtime cost is acceptable.	Same 32,768-token native DNA context, with stronger long-context behavior when YaRN is enabled.

Carbon uses one DNA token per 6 bases. A 32,768-token context is therefore about 196,608 bp before accounting for model tags and generation budget. YaRN long-context inference is available for Carbon-3B and Carbon-8B. Carbon-500M runs at its native 8,192-token context.

Settings

Setting	Default	Description
Mode	Generation	Selects generation, scoring, or comparison.
Model	Carbon-3B	Chooses Carbon-500M, Carbon-3B, or Carbon-8B.
Inference resolution	6-mer token	Uses standard 6-mer inference or Carbon's FNS base-pair marginals. Base-pair mode provides nucleotide-resolved generation and scoring with additional runtime.
Precision	bfloat16	Inference precision. bfloat16 is the standard setting. float32 can be slower and requires more GPU memory.
Species conditioning	None	Optionally adds one of Carbon's six documented taxonomic metadata tags: mammalian vertebrate, non-mammalian vertebrate, fungi, plant, protozoan, or invertebrate. FNS scoring accepts raw DNA only, so select None for base-pair score and compare jobs.
Use YaRN	Off	Enables long-context RoPE scaling for models that support it.
Max context (bp)	6144	Maximum DNA context passed to the model. Longer sequences use the rightmost window. Values are rounded to 6-mer boundaries.
Max new bp	30	Maximum generated DNA length in generation mode. The generation budget shares the model context window with the input context.
Sample	Off	Uses stochastic decoding when enabled. When off, generation is deterministic for the same model and context.
Temperature	1	Sampling temperature used only when Sample is on. Higher values increase diversity.
Top-k	50	Limits sampling to the top k candidate tokens when Sample is on.
Top-p	1	Nucleus sampling cutoff when Sample is on. Lower values concentrate generation on higher-probability choices.
Reverse-complement average	Off	In compare mode, averages forward and reverse-complement scores before calculating the delta.
Batch size	2	Number of sequences scored per inference batch in score and compare modes.
Seed	0	Random seed used only for sampled generation.

Understanding Carbon results

Generation results

Generation mode returns one row per input sequence.

Column	Meaning
input_length_bp	Length of the submitted sequence.
context_length_bp	Number of bases actually used as model context. If this is smaller than input_length_bp, the rightmost context window was used.
max_new_bp	Requested generation limit in base pairs.
generated_sequence	Newly generated DNA continuation.
generated_length_bp	Length of the generated continuation.
full_sequence	Context sequence plus generated continuation.
do_sample, temperature, top_k, top_p	Decoding settings used for the run.
inference_resolution	Whether generation used standard 6-mer token selection or FNS base-pair marginals.

The result view shows every submitted sequence and highlights each generated continuation. Downloaded generation jobs also include carbon-generated-sequences.fasta, where each record contains the context and generated continuation together.

Score results

Score mode reports log likelihoods for each sequence.

Column	Meaning
sequence_length_bp	Original sequence length.
scored_length_bp	Length actually scored after context trimming. Standard token scoring aligns to 6-mer boundaries; FNS scoring preserves every original base, including partial final 6-mers.
mean_logp	Mean log probability over scored positions or tokens. Higher values indicate a sequence the model considers more likely.
total_logp	Sum of log probabilities across the scored sequence. Total log probability becomes more negative as sequence length increases, so it should not be compared across very different lengths without care.
score_count	Number of scored units.
score_unit	6mer_token for standard inference or base_pair for FNS inference.
token_count	Backward-compatible token count for standard 6-mer scoring; empty for FNS base-pair scoring.
scoring_method	Whether Carbon used model-provided base-pair scoring or token-level log likelihood.

mean_logp is usually the most useful score for comparing sequences of similar length. It is not a calibrated biological effect size, binding score, pathogenicity probability, or expression measurement.

FNS score and compare jobs include one CSV and one JSON file per scored orientation. Each row reports the 1-based position, observed base, Carbon's A, T, C, and G marginal probabilities, the observed-base probability, and its unmodified log probability. A true zero probability is represented as -Infinity; ProteinIQ does not replace it with a numeric floor.

Because these complete per-position files are delivered with the job result, one FNS score or compare job can export at most 40,000 total scored positions. Reverse-complement comparison counts both orientations. Reduce Max context, shorten the inputs, or turn off reverse-complement averaging when a request exceeds that delivery limit. Standard 6-mer scoring retains the model context limits shown above.

Compare results

Compare mode returns one row for the reference-versus-alternate pair.

Column	Meaning
ref_mean_logp	Mean log probability for the reference sequence.
var_mean_logp	Mean log probability for the alternate sequence.
delta_mean_logp	var_mean_logp - ref_mean_logp. Positive values favor the alternate under the model.
ref_total_logp	Total log probability for the reference sequence.
var_total_logp	Total log probability for the alternate sequence.
delta_total_logp	Alternate-minus-reference total log probability.
ref_mean_logp_forward, var_mean_logp_forward, delta_mean_logp_forward	Forward-orientation component scores. When reverse-complement averaging is off, these match the main mean-log-probability columns.
ref_mean_logp_reverse_complement, var_mean_logp_reverse_complement, delta_mean_logp_reverse_complement	Reverse-complement component scores returned when reverse-complement averaging is enabled.
ref_total_logp_forward, var_total_logp_forward, delta_total_logp_forward	Forward-orientation total-log-probability component scores.
ref_total_logp_reverse_complement, var_total_logp_reverse_complement, delta_total_logp_reverse_complement	Reverse-complement total-log-probability component scores returned when reverse-complement averaging is enabled.
preferred_sequence	alternate when the alternate has higher mean log probability, otherwise reference.
rev_comp_avg	Whether forward and reverse-complement orientations were averaged.

For single-base or short edits in a fixed-length context, delta_mean_logp is the clearest comparison column. For insertions, deletions, or sequences of different length, inspect both mean and total deltas because length changes affect total log probability directly.

Downloaded files

File	Included for	Contents
carbon-results.csv	All modes	Spreadsheet-ready result rows.
carbon-results.json	All modes	Full result rows in JSON format.
carbon-generated-sequences.fasta	Generation mode	FASTA records containing the context plus generated continuation.
carbon-fns-###-*-base-probabilities.csv	FNS score and compare	Per-base A, T, C, and G marginals plus observed-base probability and log probability.
carbon-fns-###-*-base-probabilities.json	FNS score and compare	JSON form of the same per-base probability records.

Carbon examples

TP53 coding-sequence continuation with FNS

This example uses the first 240 bp of the human TP53 coding sequence from NCBI RefSeq transcript NM_000546.6 to demonstrate nucleotide-resolved Carbon generation.

• Input: NM_000546.6 TP53 coding-sequence start, 240 bp
• Non-default settings: Inference resolution = Base pair (FNS) for nucleotide-resolved generation; Max context (bp) = 240; Max new bp = 60

Carbon-3B used the full 240 bp context and returned a 60 bp continuation, highlighted in blue after the submitted sequence. The continuation is an autoregressive model output rather than a reconstruction of the known TP53 transcript, and its locally repetitive pattern should not be interpreted as evidence of gene function, expression, or biological viability.

Base-pair likelihood scoring across three bacteriophages

This batch example scores equal-length sequence starts from phiX174 (NC_001422.1), bacteriophage lambda (NC_001416.1), and M13 (NC_003287.2) in one Carbon run.

• Inputs: The first 180 bp from each of the three NCBI reference genomes
• Non-default settings: Mode = Score; Model = Carbon-500M; Inference resolution = Base pair (FNS); Max context (bp) = 180; Batch size = 3

All three sequences were scored over 180 base pairs with fns_base_pair_log_likelihood. M13 has the highest mean log probability in this run at -1.343, followed by phiX174 at -1.364 and lambda at -1.377. Because the sequences have equal length and share the same model and settings, their mean log probabilities can be compared directly here. The ordering describes Carbon's learned sequence expectation; it does not measure phage fitness, infectivity, or host range.

HBB c.20A>T reference and alternate comparison

This example compares a 120 bp coding context from human HBB transcript NM_000518.5 with the c.20A>T alternate sequence, changing the reference GAG codon to GTG.

• Inputs: Matched 120 bp reference and alternate HBB coding sequences
• Non-default settings: Mode = Compare; Model = Carbon-500M; Max context (bp) = 120; Reverse-complement average = on

After averaging the forward and reverse-complement orientations, the reference mean log probability is -5.9805, the alternate mean log probability is -6.4527, and delta_mean_logp is -0.4722. Carbon therefore marks the reference as preferred under this checkpoint and context. This likelihood difference is not a calibrated pathogenicity score and cannot replace clinical annotation or experimental evidence.

How Carbon works

Carbon is a decoder-only Transformer model family trained on DNA and RNA sequence data. The key modeling choice is its hybrid tokenizer: English and metadata tokens use a text vocabulary, while DNA inside a <dna> block uses fixed non-overlapping 6-mers. ProteinIQ handles the DNA tag internally, so submitted DNA is tokenized in Carbon's DNA mode.

The 6-mer design improves efficiency because each model token represents 6 bp. It also creates practical constraints. Input must be canonical DNA, and standard token scoring and generation align context to 6-base boundaries. Carbon trims to the rightmost usable context when an input is longer than the selected window because autoregressive generation and scoring depend on the sequence immediately before the predicted bases. FNS scoring is different: Carbon right-pads a partial 6-mer internally, then returns probabilities only for the original bases, so no submitted base is discarded.

Likelihood scoring follows the usual causal language model interpretation: the model estimates how probable each next DNA token or base is given the previous context. Higher log probability means the sequence is more expected under the selected Carbon checkpoint and context, not necessarily more functional in an experiment.

FNS base-pair mode uses Carbon's native fns model revision. It marginalizes the 4,096 possible 6-mer outputs into six nucleotide distributions for generation and scoring. This gives nucleotide-resolved likelihoods while retaining Carbon's 6-mer context efficiency. Standard 6-mer mode remains the default and is faster.

Species conditioning prepends Carbon's documented taxonomic metadata token before the DNA tag. It is optional and does not change existing unconditioned runs. It is appropriate only when the submitted context has a known broad taxonomic group; it does not identify species or validate taxonomy.

YaRN extends the rotary-position context used by the model. It is useful for long genomic contexts, but longer windows increase runtime and memory use, and very long extrapolated contexts can reduce retrieval quality. For short sequence scoring or generation, the native context is usually easier to interpret.

When to use Carbon vs alternatives

Carbon fits learned DNA sequence modeling tasks: generation, likelihood ranking, and zero-shot comparison of reference and alternate DNA sequences. It is not a multiple sequence alignment tool, variant annotation database, or wet-lab validation substitute.

Use AlphaGenome when the goal is variant effect prediction against genomic functional tracks rather than model likelihood. Use DNA to Protein Converter for translation, DNA to RNA converter for transcription, and Random DNA when a random control sequence is needed without learned genomic context.