Learning Tracks
Many tracks and dozens of sub-tracks. Every track starts with a free intro lesson.
From the transformer math to a paper at the ACL Student Research Workshop.
Four sub-tracks. Foundations first, then Fine-Tuning, Evaluation, and Publishing.
Build the transformer from first principles: the tokenization problem, training a BPE/WordPiece/SentencePiece tokenizer, embeddings and positional encoding, scaled dot-product and multi-head attention, the full block, and pretraining objectives. Ends with a from-scratch tiny transformer and a research-readiness lesson.
The adaptation lifecycle, full vs parameter-efficient fine-tuning and the memory math, LoRA and QLoRA hands-on, instruction tuning and preference optimization (DPO/RLHF), data curation and decontamination, and the compute realities of free/cheap GPUs. Ends with a measured fine-tuning experiment and a job-readiness lesson.
Why evaluation gates publication: automatic metrics (perplexity, BLEU, ROUGE, BERTScore) and their failure modes, benchmarks and data contamination, human evaluation and LLM-as-judge bias, statistical significance and ablations, and reproducibility. Ends with a reusable evaluation harness and a job-readiness lesson.
The research craft: the ACL ecosystem and ACL Rolling Review, reading papers and literature review, scoping a question and finding a mentor, reproducing a paper, writing in the ACL template (with the mandatory Limitations + Ethics sections), surviving peer review and the rebuttal, and presenting. Ends with an SRW-ready short paper and a research-career lesson.