This model is designed to provide Korean language capabilities with a focus on reasoning tasks. It is a scaled-up version of OLAIR/ko-r1-7b-v2.0.3 for experimental purposes.

2. Benchmark Performance

The model's performance has been evaluated using the HAE-RAE Reasoning Challenge (HRC), which measures reasoning abilities across five domains.

Model	Chemistry	Math	Physics	Physics Word Puzzles	Puzzles	Average
o1-2024-12-17	57.14	78.18	77.78	80.00	84.62	75.54
o3-mini-high	57.14	81.82	77.78	70.00	69.23	71.19
o3-mini-2025-01-31	50.00	80.00	70.37	50.00	76.92	65.46
o1-mini-2024-09-12	42.86	56.36	70.37	60.00	15.38	48.99
Deepseek-R1	50.00	54.55	62.96	70.00	7.69	49.04
gpt-4o-2024-11-20	35.71	32.73	51.85	50.00	53.85	44.83
Ko-R1-14B-v2.0.3	28.57	50.9	48.14	30.00	30.77	37.68
Exaone-3.5-32B-Instruct	21.43	30.91	25.93	50.00	38.46	33.35
Qwen2.5-72B-Instruct	35.71	30.91	51.85	20.00	23.08	32.31
Ko-R1-7B-v2.0.3	7.14	61.82	40.74	40.00	0.00	29.94
Ko-R1-7B-v1	7.14	63.64	37.04	40.00	0.00	29.56
gpt-4o-mini-2024-07-18	21.43	29.09	37.04	50.00	0.00	27.51
UNIVA-Bllossom_DeepSeek-llama3.1-Bllossom-8B	28.57	16.36	33.33	10.00	15.38	20.73

Comparison of Average Scores Ko-R1-14B-v2.0.3 outperforms non-reasoning models of way bigger size like Exaone-3.5-32B-Instruct, Qwen2.5-72B-Instruct, and gpt-4o-mini-2024-07-18.

Training Rewards Even when trained with the same datasets, bigger models just learn MORE.

3. Limitations

The model is still vulnerable to Korean-related inputs, leading to endless loops of thinking. We are working to fix it.

ETC

How to Cite

To be added

Contact

spthsrbwls123@yonsei.ac.kr