Instructions to use mrtineu/thinkingcap-test-submission with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use mrtineu/thinkingcap-test-submission with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="mrtineu/thinkingcap-test-submission")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("mrtineu/thinkingcap-test-submission")
model = AutoModelForCausalLM.from_pretrained("mrtineu/thinkingcap-test-submission", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use mrtineu/thinkingcap-test-submission with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "mrtineu/thinkingcap-test-submission"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mrtineu/thinkingcap-test-submission",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/mrtineu/thinkingcap-test-submission

SGLang

How to use mrtineu/thinkingcap-test-submission with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "mrtineu/thinkingcap-test-submission" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mrtineu/thinkingcap-test-submission",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "mrtineu/thinkingcap-test-submission" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mrtineu/thinkingcap-test-submission",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use mrtineu/thinkingcap-test-submission with Docker Model Runner:
```
docker model run hf.co/mrtineu/thinkingcap-test-submission
```

ThinkingCap-Test submission: Martin Senkyr

Qwen3-0.6B with one extra decoder layer inserted after layer 13. Only 64 of its MLP channels are trained (about 198k parameters), on shorter reasoning traces the student wrote with brevity hints from Qwen3-8B. It is a plain 29-layer Qwen3 checkpoint, so it loads in stock Transformers and vLLM with no custom code.

Checkpoint	Tokens saved	Accuracy change
root: extra-mlp-64, step 170 (half an epoch)	28.7% [27.5, 30.0]	−1.71 pp [−2.40, −1.02]

Scored with the unmodified evaluate.py and compare.py of ThinkingCap-Test at bd52d8d, full 8 samples per question. The method and all results are in IDEA.md and data.md.

model.safetensors, config.json, tokenizer files: the main checkpoint (evaluate.py --model mrtineu/<this repo>)
results.json: its evaluate.py output
base/results.json: the Qwen3-0.6B baseline it is compared against
base-local/results.json: the same baseline loaded by local path, the way the checkpoints were; every comparison moves by at most 0.1% tokens and 0.02 pp against it
other-checkpoints/: later steps of the same run, the same layer placed after layer 20, and the LoRA run
martin-senkyr.bundle: the code as a git bundle (git clone martin-senkyr.bundle); code/ is the same code, readable here

Downloads last month: -

Safetensors

Model size

0.6B params

Tensor type

BF16

Model tree for mrtineu/thinkingcap-test-submission

Base model

Qwen/Qwen3-0.6B-Base

Finetuned

Qwen/Qwen3-0.6B

Finetuned

(1438)

this model

mrtineu
/

thinkingcap-test-submission

ThinkingCap-Test submission: Martin Senkyr

Contents

Model tree for mrtineu/thinkingcap-test-submission