Replication with Purpose

Resources

• Paper

  Learning Representations by Back-Propagating Errors (1986)

  Foundational reference for gradient-based learning; a core anchor when reproducing modern optimization behavior from first principles.

• Paper

  Attention Is All You Need

  Baseline transformer replication target with clear architecture and measurable points of divergence under modern tooling choices.

• Paper

  Language Models are Few-Shot Learners

  Useful for evaluating scaling-era claims and benchmark sensitivity when training setup, data curation, or prompt protocols shift.

• Repository

  minGPT

  Compact transformer codebase ideal for controlled replications and targeted ablations without hidden framework complexity.

• Guide

  Papers With Code

  Reference map for implementation baselines, benchmark status, and available repos before selecting replication targets.

Writing

Template | Apr 2026

Replication Log Template

Use this structure for both external writeups and full on-page replication logs with plots, code, and error analysis.

Goal and Claim

State the exact claim being replicated and define success metrics before running experiments.

python train.py --dataset wikitext --layers 12 --heads 12 --seed 17

Videos