Evo 2

How SubSeq Runs It

• Jobs run /ref/bin/evo2_subseq.py with the Evo 2 runtime.
• The default checkpoint is evo2_7b. Advanced jobs can choose evo2_7b_262k or evo2_7b_base.
• Model data is loaded from the service-local Evo 2 ref tree.
• Network access is disabled for jobs, and output is forced to /outputs.

Input

Use typed DNA sequence text, one FASTA file, or a folder of FASTA files. Sequences should use A, C, G, T, and N.

In the job form, choose the operation first, then choose Sequence Input. If you choose a FASTA file or folder, the next source selector only asks where that file or folder comes from: paste/upload, dataset, or previous job output.

>example
ACGTACGTACGTACGT

Example Arguments

Generate

generate
--sequence=ACGTACGTACGT
--name=prompt
--n-tokens=400
--num-samples=1
--temperature=1.0
--top-k=4

Score

score
--fasta=/inputs/sequences.fasta
--reduce-method=mean
--average-reverse-complement

Embeddings

embed
--input-folder=/inputs/fastas
--recursive
--pool=mean

Arguments

Global

• --model-name=evo2_7b|evo2_7b_262k|evo2_7b_base: choose the Evo 2 7B-family checkpoint.

Input

• --sequence=<dna>: one typed DNA sequence.
• --name=<label>: output label for typed sequence input.
• --fasta=/inputs/example.fasta: one FASTA file.
• --input-folder=/inputs/fastas: folder of FASTA files.
• --recursive: search subfolders when using a FASTA folder.

Generation

• --n-tokens=<int>: number of new DNA tokens. Default: 400.
• --num-samples=<int>: continuations per prompt. Default: 1.
• --temperature=<float>: sampling temperature. Default: 1.0.
• --top-k=<int>: top-K sampling cutoff. Default: 4.
• --top-p=<float>: nucleus sampling cutoff. Default: 1.0.

Scoring

• --batch-size=<int>: scoring batch size. Default: 1.
• --reduce-method=mean|sum: length-normalized or total score. Default: mean.
• --average-reverse-complement: average each score with the reverse-complement score.
• --prepend-bos: prepend Evo 2's beginning-of-sequence token when scoring.

Embeddings

• --layer-name=<name>: Evo 2 layer to extract. Default: blocks.28.mlp.l3.
• --pool=mean|tokens|both: write one vector per sequence, per-token vectors, or both.

Outputs

• Generation: /outputs/generated.fasta and /outputs/generated.json.
• Scoring: /outputs/scores.tsv and /outputs/scores.json.
• Embeddings: /outputs/embeddings_manifest.json and one or more .npy arrays.

Submit

Queue a run from New Job -> Evo 2.