EPFL-TA-Meister — LLM as a Teaching Assistant
Llama-3-8B fine-tuned as an EPFL teaching assistant: SFT, DPO alignment, RAG over coursebooks, then 4-bit/8-bit/GPTQ/AWQ quantization. Five models shipped to HuggingFace.
What this is
Llama-3-8B fine-tuned as a teaching assistant for EPFL coursework. Three milestones: literature review and preference data, then SFT + DPO, then RAG + quantization.
How it works
M2: SFT on a preference dataset built from EPFL course material, then DPO on top to align answer style with what students prefer.
M3: RAG over ~5 GB of EPFL coursebooks for grounded answers. Quantized the resulting checkpoints four ways: 4-bit (bitsandbytes), 8-bit, GPTQ-4bit, AWQ-4bit. Benchmarked each for quality vs latency.
Results
Five models on HuggingFace: PeterAM4/EPFL-TA-Meister, -SFT, -GPTQ-4bit, -4bit, -AWQ-4bit. All three milestone repos public: mnlp_p1 (lit review + preference data), mnlp_p2 (SFT + DPO), mnlp-p3 (RAG + quantization).
CS-552 Modern NLP at EPFL, team “4-pack”.