werty1248/Qwen2.5-32B-s1.1-Ko-Native

Korean S1K-1.1

• Generation Result: werty1248/qwen-32b-s1.1-Ko-Native-result
  • AIME2024(Korean): 30% (9/30) with max_tokens=8192
• Chinese appears in think tokens (Not halucination or failure behavior. Solved correctly in Chinese, then answered in Korean)

Training Details

• Train with s1k official code
  • block_size = 20000
  • gradient_checkpointing=True
• Training Dataset: werty1248/s1k-1.1-Ko-ReGenerated-Formatted
  • Translated questions from s1k-1.1 into Korean, and then regenearted using Deepseek-R1.
  • Need to modify "text" column to Qwen format
• 8x H200 SXM, 2 hours; $63.84, except for the cost of trial and error :(
• 

Evaluation

• HRM8K - Korean Math Benchmark

Generated and evaluated with my own code; accuracy may differ.

Model	GSM8K	KSM	MATH	OMNI_MATH
Qwen2.5-32B-s1.1-Ko-Native	89.92	39.85	87.73	42.06
*GPT-4o	91.21	22.83	74.45	30.75
*GPT-4o-mini	87.57	19.40	70.68	26.45
EXAONE-3.5-7.8B-Stratos-Ko	83.02	15.97	67.49	24.62
Qwen2.5-7B-s1.1-Ko-Native	76.27	15.48	66.45	23.57
EXAONE-3.5-7.8B-Instruct	81.58	14.71	63.50	21.69
*Qwen2.5-14B-Instruct	66.34	15.55	53.38	20.64
*Llama-3.1-8B-Instruct	77.79	7.21	49.01	15.92
*Qwen2.5-7B-Instruct	58.38	13.10	48.04	16.55
*EXAONE-3.0-7.8B-Instruct	72.33	7.98	46.79	15.35
*Ko-R1-1.5B-preview	43.3	?	73.1	29.8

* Reported by HRM8K authors

• Generation
  • temperature = 0.7
  • top_p = 0.95
  • max_tokens = 8192
  • If it exceeded the maximum tokens but did not generate </think> tokens, add </think> tokens and generate 512 additional tokens
• Evaluation
  • Custom parser & evaluation code; may have parsed incorrectly

---

Why Qwen? Why EXAONE can't?