nectec/pathumma-thaillm-8b-think-3.0.0

Pathumma-ThaiLLM-Think-3.0.0 Overview

Pathumma-ThaiLLM-Think-3.0.0 is a post-trained Thai Large Language Model from NECTEC, developed on the Thai national initiative's ThaiLLM foundation model. It undergoes a two-stage Supervised Fine-Tuning (SFT) process to significantly enhance its capabilities.

Key Capabilities

• Instruction Following: Improved compliance with user instructions.
• Tool/Function Calling: Enhanced ability to handle structured tool and function calls.
• Mathematical & Coding Competence: Specialized in mathematical problem-solving, code generation, and analysis.
• Multi-step Analytical Reasoning: Designed for complex, multi-step analytical tasks.
• Bilingual Robustness: Strong performance in both Thai and English.

Training Strategy

The model's post-training involves two distinct stages:

1. Instruction & Tool-Calling Alignment: Focuses on instruction compliance, structured tool-call formatting, general Thai task robustness, and STEM-oriented instruction alignment using datasets like beyoru/ToolCall_synthetic_qwen3 and airesearch/WangchanX-FLAN-v6.
2. Reasoning Specialization: Concentrates on multi-step mathematical analysis, code understanding and synthesis, structured analytical responses, and tool-calling with explicit reasoning traces, utilizing datasets such as nvidia/OpenMathReasoning and nvidia/OpenCodeReasoning.

Training was conducted on the LANTA high-performance computing cluster, leveraging 64 A100 40GB GPUs. The model aims to advance sovereign Thai large language models optimized for analytical and tool-augmented intelligence.

Limitations

Users should be aware that the model may hallucinate if the tool schema is incomplete, and performance on long analytical chains might degrade without retrieval mechanisms. Its domain coverage is dependent on the included training corpora.