Adapter format#

The only weights on the wire in protocol v1 are a LoRA adapter for Qwen3-0.6B-Base. The layout is fixed.

Property	Value
Base model	`Qwen/Qwen3-0.6B-Base` (28 layers, hidden size 1024)
Method	LoRA (PEFT), `r = 16`, `alpha = 32`, dropout 0
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj` in every layer
File format	safetensors
dtype	float32 (`F32`)
Tensors	392 (28 layers × 7 modules × A and B)
Size	≈ 10.09 M parameters / ≈ 40.4 MB
Metadata	none

Tensor names#

base_model.model.model.layers.{L}.self_attn.{q|k|v|o}_proj.lora_{A|B}.default.weight
base_model.model.model.layers.{L}.mlp.{gate|up|down}_proj.lora_{A|B}.default.weight

L = 0…27.

Shapes#

lora_A is [16, in], lora_B is [out, 16]:

Module	`lora_A`	`lora_B`
`q_proj`	`[16, 1024]`	`[2048, 16]`
`k_proj`, `v_proj`	`[16, 1024]`	`[1024, 16]`
`o_proj`	`[16, 2048]`	`[1024, 16]`
`gate_proj`, `up_proj`	`[16, 1024]`	`[3072, 16]`
`down_proj`	`[16, 3072]`	`[1024, 16]`

Validation (coordinator side)#

Reject with 422 unless the upload has exactly the same key set as the global adapter, identical shape and dtype per tensor, only finite values, and is valid safetensors (pickle or other formats are rejected).

The safetensors header is plain JSON, so you can inspect a file without PyTorch:

import json, struct
f = open("tur-3.safetensors", "rb")
n = struct.unpack("<Q", f.read(8))[0]
baslik = json.loads(f.read(n))
adlar = [k for k in baslik if k != "__metadata__"]
print(len(adlar))   # 392