Dataset profile
Most sequences fit comfortably, but the tail is long
| Training rows | 6,566 |
| Test rows | 119 |
| Training median length | 308 |
| Training mean length | 435.55 |
| Training maximum | 7,540 |
| Training rows ≤512 | 4,960 |
02
Each protein becomes a fixed-width sequence of residue labels
Data and objectiveBottom lineThe input is a variable-length amino-acid sequence; the target is a parallel Q8 label sequence. Padding standardizes the tensor shape, and the custom metric removes padded positions from evaluation.
Input
Amino-acid sequence
Each row contains a protein sequence and its residue-level expected labels.
→
Encode
Tokenize n-grams
The reusable code supports n-grams; the executed `main.py` sets n = 1.
→
Shape
Pad or truncate
Short sequences are post-padded; long sequences are pre-truncated to 512 positions.
→
Target
Classify each residue
A nine-way softmax represents eight Q8 states plus padding.
Residue-label mix
Helices and strands dominate the labeled positions
H
32.84%
E
22.58%
–
19.54%
T
11.16%
S
8.19%
G
3.96%
B
1.16%
I
0.57%