Multi-agent AI Teaching Assistant training system that learns from limited course materials. It provides an end-to-end pipeline — dataset generation, QLoRA fine-tuning, and RAG augmentation — to produce domain-specific TA models with full local deployment support.
InternTA is a multi-agent automated AI Teaching Assistant training system designed for courses with scarce teaching materials, validated in synthetic biology education. The system operates through three collaborative agents:
Dataset Agent: Extracts exercises, terms, and concepts from Excel course materials, generating OpenAI conversation-format training data with explicit reasoning paths and guided answering strategies for thought-provoking questions, outputting training.json / validation.json.
Training Agent: Uses DeepSeek-R1-Distill-Qwen-7B as the base model, performing efficient fine-tuning via PEFT + QLoRA (4-bit quantization), with a built-in LLM Judge for automated training plan generation and hyperparameter tuning, supporting both basic SFT (train.sh) and advanced Agent-driven training (traino.sh) modes.
RAG Agent: Performs structured processing and semantic retrieval over course materials during inference, injecting relevant knowledge fragments into the generation process to compensate for fine-grained knowledge gaps in the fine-tuned model.
End-to-End Data Flow#
Excel raw data → data/generate_data.py → training.json / validation.json
↓
train_agent.py or sft_internTA2.py (QLoRA fine-tuning)
↓
merge.py (merge LoRA adapter)
↓
api.py (FastAPI) + app.py (Streamlit) → User access
Deployment & Integration#
- Full local deployment on 8GB+ VRAM GPU, preventing data leakage
- OpenAI-compatible
/v1/chat/completionsendpoint with Bearer Token authentication - Dual entry: Streamlit Web interface (default port 8080) + FastAPI API service
Dockerfile.webanddocker-compose.web.ymlexist; usage details TBD
Quick Start#
git clone https://github.com/kongfoo-ai/internTA
cd internTA
pip install -r requirements.txt
sh run.sh
Unconfirmed Information#
- Associated paper: Not directly referenced in README; may exist but unlisted
- Online demo: "E. Copi (Education)" mentioned but no specific URL provided
- Model weights/checkpoints: Not published in the repository
- Quantitative evaluation results: Described as strong but no specific metrics disclosed
- RAG implementation details: Vector database/embedding model not specified
- LLM Judge specifics: Evaluation criteria and dependent model not detailed