LEADERBOARD.md

🏆 Leaderboard

Subtask 1 — Term Detection

Identify technical term boundaries in EN-TR sentence pairs.

System	Type	P	R	F1
GPT-5.2	baseline	0.82	0.92	0.87
Claude Sonnet 4.5	baseline	0.71	0.81	0.75
Gemini 3 Flash	baseline	0.56	0.80	0.66
DeepSeek v3.2	baseline	0.51	0.70	0.59
Co-Text · Mistral Large	participant	0.45	0.71	0.55
Co-Text · Mistral Large + Para	participant	0.44	0.70	0.54
Koç-CoT · Llama 4 Scout	participant	0.42	0.50	0.46
Co-Text · Llama 4 Scout	participant	0.42	0.49	0.46
Llama 4 Scout	baseline	0.42	0.51	0.46
KU-RAG · Llama 4 Scout	participant	0.41	0.50	0.45
Co-Text · Llama 4 Scout + Para	participant	0.42	0.44	0.43
Human Expert	human	—	—	0.83

Subtask 2 — Term Correction with Expert Input

Fix term translations using expert hints. Evaluated by Exact Match (EM).

System	Type	EM
GPT-5.2	baseline	0.40
Claude Sonnet 4.5	baseline	0.40
DeepSeek v3.2	baseline	0.38
Co-Text · Mistral Large	participant	0.34
Co-Text · Mistral Large + Para	participant	0.34
Koç-CoT · Llama 4 Scout	participant	0.33
Co-Text · Llama 4 Scout	participant	0.30
Gemini 3 Flash	baseline	0.30
KU-RAG · Llama 4 Scout	participant	0.27
Llama 4 Scout	baseline	0.24
Co-Text · Llama 4 Scout + Para	participant	0.24
Human Expert	human	0.60

Subtask 3 — End-to-End Post-Edit

Full translation post-editing with offline glossary access.

System	Type	chrF	BLEU
New Mind AI · Qwen3-4B	participant	85.79	70.27
Co-Text · Llama 4 Scout + Para	participant	85.64	69.42
GPT-5.2	baseline	83.97	64.82
Co-Text · Llama 4 Scout	participant	82.37	62.53
Llama 4 Scout	baseline	81.70	62.59
Claude Sonnet 4.5	baseline	80.43	57.93
Koç-CoT · Llama 4 Scout	participant	79.77	58.27
DeepSeek v3.2	baseline	78.83	55.18
Co-Text · Mistral Large + Para	participant	77.68	50.39
Co-Text · Mistral Large	participant	76.67	49.33
Gemini 3 Flash	baseline	74.59	52.82
KU-RAG · Llama 4 Scout	participant	66.78	36.38