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