Instructions to use qweqwqw113/cerpt-korean-stage1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use qweqwqw113/cerpt-korean-stage1 with Transformers:

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

pipe = pipeline("text-generation", model="qweqwqw113/cerpt-korean-stage1")

# Load model directly
from transformers import AutoModelForSeq2SeqLM
model = AutoModelForSeq2SeqLM.from_pretrained("qweqwqw113/cerpt-korean-stage1", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use qweqwqw113/cerpt-korean-stage1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "qweqwqw113/cerpt-korean-stage1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "qweqwqw113/cerpt-korean-stage1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/qweqwqw113/cerpt-korean-stage1

SGLang

How to use qweqwqw113/cerpt-korean-stage1 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 "qweqwqw113/cerpt-korean-stage1" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "qweqwqw113/cerpt-korean-stage1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "qweqwqw113/cerpt-korean-stage1" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "qweqwqw113/cerpt-korean-stage1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use qweqwqw113/cerpt-korean-stage1 with Docker Model Runner:
```
docker model run hf.co/qweqwqw113/cerpt-korean-stage1
```

CERPT Korean Stage 1

CERPT (Causal Evidence Recursive Program Transformer) is a research architecture that keeps typed intermediate state in a workspace and updates that state through explicit operators and evidence checks.

This checkpoint is a small Korean Stage 1 experiment. It is trained from scratch on Korean math, logic, and multi-hop reasoning examples converted to four cycles:

EXTRACT → SIMULATE/BIND → CHECK → WRITE_RESULT

Status

This is not a general-purpose LLM. It is a research checkpoint for testing whether a small CERPT model can learn Korean task solving and structured reasoning traces. It should not be used as an open-domain assistant or for high-stakes decisions.

Training data

OpenMathReasoning-mini-ko, CC-BY-4.0
HRMCR, Apache-2.0
Ko-StrategyQA, Apache-2.0

The local conversion and split details are documented in docs/KOREAN_STAGE1_DATA.md. Dataset providers and licenses must be credited when redistributing derived artifacts.

Architecture

PyTorch and Hugging Face compatible CERPTForConditionalGeneration
4 recursive workspace cycles
6 typed operators: EXTRACT, BIND, COMPARE, SIMULATE, CHECK, WRITE_RESULT
Korean tokenizer trained from the Stage 1 text

Limitations

No general web or code pretraining
No reliable image/video understanding training in this checkpoint
Structured output quality and Korean open-domain ability are not established
Reported loss is a training diagnostic, not a general capability score

Downloads last month: -

Safetensors

Model size

15.9M params

Tensor type

F32