Name
A string. A .npy file has one array and no name. A container such as safetensors or GGUF gives every tensor a name, often with dots for the layer it came from.
One tensor
The array inside the file
Every container on this site is a different wrapper around the same idea: one named array, stored so another program can read it back.
Dtype
Which kind of number is stored: float32, bfloat16, an integer, or a block of quantized weights. The file’s own name for that dtype is usually different. The dtype table lines them up.
Shape
The length of each dimension, outermost first. A matrix NumPy would call (rows, cols) is written that way here. GGUF lists dimensions in the opposite order, so a reader reverses them.
Bytes
Little-endian, packed with no gaps, one row after another. Narrow types such as 4-bit values share a byte. Quantized block types store a fixed-size block along the innermost dimension.
A container is the file around those tensors: a header, the names, and usually a little metadata. See the containers.
One set of arrays
The same arrays in the specs test vector. A file is listed only when that array was read back out of it.
| Name | Dtype | Shape | Read back from |
|---|---|---|---|
| w_f32 | float32 | 4 × 64 | safetensors, npz, npy, gguf, raw-sidecar |
| b_f16 | float16 | 8 | safetensors, npz, gguf, raw-sidecar |
| e_bf16 | bfloat16 | 2 × 3 | safetensors, npz, gguf, raw-sidecar |
| i_i32 | int32 | 3 × 2 | safetensors, npz, gguf, raw-sidecar, text-csv-tsv-json |
| s_f32 | float32 | [] | safetensors, npz, raw-sidecar |