Teaching LLMs to think before they teach

What it does

EduChat is a vertical-domain chat model built by East China Normal University for teachers, students, and parents. It handles typical classroom tasks—automatic question generation, essay correction, and course tutoring—but also layers in emotional support and mental-health guidance. The project ships fine-tuned weights atop Qwen and Baichuan, a roughly four-million-sample mixed instruction dataset, and a data-cleaning utility called CleanTool.

The interesting bit

The latest release, EduChat-R1, adopts a “Thinking before teaching” paradigm: it uses structured chain-of-thought reasoning tailored for pedagogy and claims to blend multiple psychological theories for counseling scenarios. It also supports the MCP framework for calling external tools, which nudges the model from simple chatbot toward agentic teaching assistant.

Key highlights

• Model sizes span 1.8B to 32B parameters, with recent versions based on Qwen3.0 and Qwen2.5.
• Niche educational scenarios: Socratic heuristic teaching, values alignment, and “mind-care” counseling.
• Open-sources both model weights and a mixed Chinese-English instruction dataset of about 4 million samples.
• Includes CleanTool, a data-cleaning pipeline with optional GPU-accelerated deduplication and low-quality filtering.
• Live web demo runs at educhat.top.

Caveats

• The README and system prompts are entirely in Chinese, so English-only readers will need translation help.
• No quantitative benchmarks or evaluation numbers are provided; performance claims are purely qualitative.
• The project has cycled through four base-model families in two years, so older releases feel more like historical artifacts than active branches.

Verdict

Worth a look if you research Chinese-language educational AI or need a ready-made tutoring backbone. Skip it if you are hunting for state-of-the-art general reasoning or fully English-native documentation.