ubitus/Lilith-70B-Instruct

Overview

This model is a fine-tuned version of LLaMA 3.3 70B, optimized for multilingual benchmarks including TMMlu+, TMlu, and MMLU. The fine-tuning process focused on enhancing reasoning, comprehension, and domain-specific performance. This model was developed as part of an iterative pipeline leveraging large-scale datasets and Chain-of-Thought (CoT) methodologies.

---

Key Features

•	Base Model: LLaMA 3.3 70B
•	Dataset Sources: Custom-generated using LLMs, focused on high-quality, multilingual tasks.
•	Chain-of-Thought Fine-Tuning: Enhanced logical reasoning with curated datasets.

Data Preparation

1.	Custom Dataset Generation
2.	Traditional Chinese Data Filtering

Evaluation

Please checkout Open TW LLM Leaderboard for full and updated list.

Model	TMMLU+	TMLU	Function Calling
ubitus/Lilith-70B-Instruct	76.06%	73.70%	✅
Llama-3-Taiwan-70B-Instruct	67.53%	74.76%	✅
Qwen1.5-110B-Chat	65.81%	75.69%	✅
Yi-34B-Chat	64.10%	73.59%	✅
Meta-Llama-3-70B-Instruct	62.75%	70.95%	✅
Llama-3-Taiwan-8B-Instruct	52.28%	59.50%	✅
Mixtral-8x22B-Instruct-v0.1	52.16%	55.57%	✅
Gemini-1.5-Pro	49.92%^	61.40% (5-shot)	✅
Breexe-8x7B-Instruct-v0_1	48.92%	-	❓
Breeze-7B-Instruct-v1_0	41.77%	55.57%	❓
Llama3-TAIDE-LX-8B-Chat-Alpha1	39.03%	47.30%	❓
Claude-3-Opus	-	73.59% (5-shot)	✅
GPT4-o	-	65.56% (0-shot), 69.88% (5-shot)	✅

This model is well-suited for:

1.	Multilingual Comprehension Tasks: Designed to handle diverse languages and formats.
2.	Domain-Specific Applications: Excels in logical reasoning and structured problem-solving.
3.	Benchmarks and Testing: An excellent choice for academic and industrial evaluations in multilingual NLP.