Native Tinker verbalization-training adapters
This repository archives three rank-32 LoRA adapters used in Training LLMs to Verbalize Evaluation Awareness.
Each adapter is stored on its own branch because native Tinker checkpoint archives use the same filenames: adapter_model.safetensors and adapter_config.json.
| Branch | Base model | Training factor | Native size |
|---|---|---|---|
| Qwen3.6-35B-A3B-VT-adapter | Qwen/Qwen3.6-35B-A3B | F8 | 2.1 GB |
| Kimi-K2.6-VT-adapter | moonshotai/Kimi-K2.6 | F2 | 17.4 GB |
| Inkling-VT-adapter | thinkingmachines/Inkling | F2 | 18.8 GB |
Public Tinker checkpoints
The final VT adapters used in the paper's three main model comparisons are also public on Tinker. You can load them directly without downloading or converting the native adapter files.
| Model | Public checkpoint URI |
|---|---|
| Qwen3.6-35B-A3B | tinker://a2a724e5-7b3e-50db-89c7-211e334c15c5:train:0/sampler_weights/step_250 |
| Inkling | tinker://f65e7774-d0dd-5780-891c-a785cb892eb1:train:0/sampler_weights/step_200 |
| Kimi K2.6 | tinker://dbef440e-6976-5b1e-93b7-764672f6c2b5:train:0/sampler_weights/step_100 |
These are the full-method VT adapters, not ablations or SDF compositions. The Inkling checkpoint is the primary, single-round model. All three currently expire on January 1, 2028 (UTC).
Set your own TINKER_API_KEY, then load a checkpoint:
import tinker
service = tinker.ServiceClient()
sampler = service.create_sampling_client(
model_path="tinker://a2a724e5-7b3e-50db-89c7-211e334c15c5:train:0/sampler_weights/step_250"
)
Replace model_path with any URI in the table. When generating, use the matching base model's tokenizer and chat renderer. See the Tinker sampling documentation for constructing sampling requests.
Format
These are native Tinker LoRA checkpoint archives. They are not guaranteed to load directly in standard PEFT or vLLM tooling because MoE adapters retain grouped 3D expert tensors.
Convert to PEFT
Download the desired branch and obtain the matching base-model revision locally. Then run:
from tinker_cookbook.weights import build_lora_adapter
build_lora_adapter(
base_model="/path/to/base-model",
adapter_path="/path/to/native-adapter",
output_path="/path/to/peft-adapter",
)
Run fixed-prompt parity checks before using the converted adapter.