THaLLE-0.2-ThaiLLM-8B-fa: Financial and Thai Language LLM

This 8 billion parameter model, developed by KBTG Labs, is an extension of ThaiLLM-8B, specifically designed to enhance instruction-following and integrate specialized financial knowledge. It was constructed using mergekit to combine ThaiLLM-8B with Qwen3-8B and THaLLE, a component trained on 80 CFA examination sets.

Key Capabilities

Domain-Specific Financial Knowledge: Incorporates deep financial understanding derived from THaLLE's fine-tuning on CFA examination data.
Enhanced Thai Language Understanding: Builds upon ThaiLLM-8B to offer improved comprehension and generation in the Thai language.
Flexible Thinking Modes: Supports switching between "thinking" and "non-thinking" modes, similar to Qwen3-8B, which can influence performance on complex tasks.
Instruction Following: Designed for enhanced instruction-following capabilities.

Performance Highlights

Evaluations show THaLLE-0.2-ThaiLLM-8B-fa achieving strong results, particularly in "Thinking" mode. It outperforms Qwen3-8B and ThaiLLM-8B-Instruct on several financial and general knowledge benchmarks:

M3 Exam: Achieved 0.779 in Thinking mode (vs. 0.706 for Qwen3-8B and 0.720 for ThaiLLM-8B-Instruct).
M6 Exam: Achieved 0.678 in Thinking mode (vs. 0.590 for Qwen3-8B and 0.661 for ThaiLLM-8B-Instruct).
Flare CFA*: Achieved 0.852 in Thinking mode (vs. 0.806 for Qwen3-8B and 0.820 for ThaiLLM-8B-Instruct).

Good for

Applications requiring strong performance in Thai language processing.
Use cases demanding specialized financial knowledge and reasoning.
Scenarios where instruction-following and contextual understanding are critical.
Exploring the impact of different reasoning modes on model output.