Skip to content

Latest commit

 

History

History
235 lines (165 loc) · 7.1 KB

DEVELOPMENT.md

File metadata and controls

235 lines (165 loc) · 7.1 KB

🧠 ReplicateAI Development Notes

This document records all development operations, workflows, and progress logs for ReplicateAI.
It serves as both a quick-start guide and an internal research diary.

⚙️ 1. Environment Setup

Before running any commands, make sure your Python environment is ready.

uv sync
source .venv/bin/activate
pip install -r requirements.txt   # optional, if you have dependencies later

All Makefile commands assume Python is in .venv/bin/python.

🧩 2. Initialize the Project Structure

Run this once after cloning the repository:

make init

This will create the following directory layout:

ReplicateAI/
├── stage1_foundation/
├── stage2_representation/
├── stage3_deep_renaissance/
├── stage4_statistical/
├── stage5_neural_origins/
├── scripts/
├── paper_template/
└── PAPER_INDEX.json

Each stage corresponds to a historical era of AI research:

Stage Era Description
stage1_foundation 🪐 Modern Foundation LLMs & Multimodal Models (2023–2025)
stage2_representation 🔍 Representation Transformers, BERT, Embeddings
stage3_deep_renaissance 🧩 Deep Renaissance CNNs, Autoencoders
stage4_statistical 📊 Statistical SVMs, Random Forests, EM
stage5_neural_origins 🧬 Neural Origins Perceptron, Backprop, Hopfield

➕ 3. Add a New Paper Module

Use the make add command to create a new paper entry.

Example:

make add name="Qwen2.5" year=2025 org="Alibaba" stage=foundation
make add name="BERT" year=2018 org="Google" stage=representation
make add name="Perceptron" year=1958 org="Cornell" stage=neural

Each command will:

  • Create a folder like stage1_foundation/2025_Qwen2.5/
  • Copy template files from paper_template/
  • Add metadata to PAPER_INDEX.json

After running, check that new directories and index entries were created.

📊 4. View Project Progress

Check current reproduction progress by stage:

make status

Example output:

📊 Current Progress by Stage:
🪐 Modern Foundation       :   1 papers
🔍 Representation           :   1 papers
🧩 Deep Renaissance         :   0 papers
📊 Statistical              :   0 papers
🧬 Neural Origins           :   1 papers

📄 5. List All Papers

To see all registered papers from PAPER_INDEX.json:

make list

Example output:

2025 | Qwen2.5 | Alibaba | planned
2018 | BERT | Google | planned
1958 | Perceptron | Cornell | planned

🧮 6. Generate a Summary Report

To quickly summarize total paper count and stage distribution:

make report

Example output:

📅 Report generated: 2025-10-18
🧩 Total Papers: 3
  Foundation      -> 1 papers
  Representation  -> 1 papers
  Neural Origins  -> 1 papers

🧹 7. Clean Temporary Files

When you want to clean Python caches and temp files:

make clean

This removes:

  • __pycache__/
  • .pyc files

🧱 8. Implement the Paper Code

Each paper module created by make add contains a standard structure:

2025_Qwen2.5/
├── README.md          ← Paper summary & key ideas
├── report.md          ← Experiment results / analysis
├── notebook/          ← Interactive notebooks
├── src/               ← Implementation code
└── references.bib     ← Original citation

Step-by-step:

  1. Open the new folder (e.g. stage1_foundation/2025_Qwen2.5/)
  2. Edit README.md following the paper_template
  3. Implement your model under src/
  4. Add experiment results in report.md
  5. Update status to "in progress" or "completed" in PAPER_INDEX.json

🧾 9. Commit and Push

git add .
git commit -m "Add Qwen2.5 reproduction module"
git push

🧭 10. Quick Reference Cheat Sheet

Task Command
Initialize repo make init
Add paper make add name="..." year=YYYY org="..." stage=<stage>
Check status make status
List all papers make list
Generate summary report make report
Clean temp files make clean

🧠 11. Development Log (Timeline)

Date Update Notes
2025-10-18 ✅ Initialized ReplicateAI structure Added Makefile, scripts, and template
2025-10-18 ➕ Added BERT Verified indexing and make status output
2025-10-18 ✅ Attention All You Need Implemented Scaled Dot-Product Attention, MultiHeadAttention and PositionwiseFeedforward
2025-10-18 ✅ Attention All You Need Implemented Encoder, Decoder, Transformer,add a toy dataset for train test
2025-10-21 ✅ Attention All You Need Implemented training on multi30k dataset training
2025-10-22 ✅ Attention All You Need Debug and fixed loss NAN issue, and refacted the code
2025-10-23 ✅ Attention All You Need Implemented multiple tokenizer to abalition test

You can continue appending to this table as the project evolves.

🧭 12. Stage Naming Summary

Stage Code Directory Display Name Era
foundation stage1_foundation 🪐 Modern Foundation 2023–2025
representation stage2_representation 🔍 Representation 2013–2020
deep stage3_deep_renaissance 🧩 Deep Renaissance 2006–2014
statistical stage4_statistical 📊 Statistical 1990s–2000s
neural stage5_neural_origins 🧬 Neural Origins 1950s–1980s

💬 Notes

  • Always verify that PAPER_INDEX.json stays valid JSON.
  • Use make help to recall all supported commands.
  • Each reproduction module should be independent and documented.
  • When in doubt: check paper_template/README.md.

🧩 ReplicateAI — Rebuilding AI, one paper at a time.