File size: 2,753 Bytes
bb04ca4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3346de4
bb04ca4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3346de4
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
---
license: cc-by-4.0
pretty_name: CodonTranslator Data
task_categories:
  - text-generation
tags:
  - biology
  - dna
  - codon-optimization
  - protein-conditioned-generation
size_categories:
  - 10M<n<100M
---

# CodonTranslator Data

This repository contains the final public training-data release used for CodonTranslator.

## Contents

- `train/`: representative-only training shards
- `val/`: representative-only validation shards
- `test/`: representative-only held-out test shards
- `embeddings_v2/`: precomputed species conditioning embeddings used in training
- `_work/final_representative_counts.json`: final released split sizes
- `_work/split_report.json`: split audit report
- `_work/mmseqs_manifest.json`: MMseqs version and clustering parameters
- `_work/cluster_split.parquet`: cluster-level split assignments
- `_work/seq_cluster.parquet`: MMseqs cluster assignments
- `_work/seq_split.parquet`: split assignments before representative selection

## Split definition

The public `data_v3` split was rebuilt from `data_v2` with the following rules:

- MMseqs clustering in **protein space**
- test holdout by **binomial species**
- validation split from **seen species but unseen clusters**
- representative-only parquet outputs, one retained row per released representative sequence

Mixed seen/held-out protein clusters and exact-protein leakage cases are removed on the seen side before final release.

## Final released split sizes

- `train = 36,888,301`
- `val = 373,637`
- `test = 331,455`

All three released splits satisfy:

- exact protein overlap `train/val = 0`
- exact protein overlap `train/test = 0`
- test species not seen in train/val
- representatives-only rows, so `rows == unique_seq_id`

## Parquet schema

Each released parquet shard contains these columns:

- `RefseqID`
- `protein_refseq_id`
- `protein_seq`
- `cds_DNA`
- `taxon`
- `shard`

## Embeddings

`embeddings_v2/` is the exact species embedding store used in training. It contains:

- `species_vocab.json`
- `species_index.json`
- `species_tok_emb.bin`
- `metadata.json`
- `taxonomy_database.json`

The released `data_v3` taxa are fully covered by this embedding store.

## Model and code

The corresponding public model and code release is:

- `alegendaryfish/CodonTranslator`

## Attribution

This release redistributes processed training data and derived embeddings for reproducibility of the CodonTranslator experiments. See the audit files in `_work/` for construction details and final verification outputs.

The raw MMseqs working database is not included because it is large, machine-specific intermediate state. The released `_work/` files are the reproducibility artifacts needed to audit and reconstruct the clustering and split decisions.