AUTORESEARCH

AUTORESEARCH Hết Thời Thức Đêm. Thời của kỹ sư thức đêm canh training, kết thúc rồi. Thức Trống Canh Train Kỷ Nguyên Kết Thúc TONG TAI AI - @tongtaiai Canh Loss Curve. Chỉnh depth, batch size, bấm chạy rồi ngồi canh loss curve. Chỉnh Depth, Batch Size Ngồi Canh Loss Curve TONG TAI AI - @tongtaiai Crash Lúc Nửa Đêm. Ngủ quên, sáng dậy run đã crash từ 2 tiếng trước. Ngủ Quên Lúc Nửa Đêm Run Crash 2 Tiếng TONG TAI AI - @tongtaiai Không Tại Bạn Đâu. Việc lặp lại hàng trăm lần là việc của máy, không phải người. Việc Lặp Lại, Máy Móc Không Phải Lỗi Của Bạn TONG TAI AI - @tongtaiai 92 Nghìn Sao, 3 Tuần. Andrej Karpathy tung ra autoresearch — 92.297 sao chỉ sau 3 tuần. About AI agents running research on single-GPU nanochat training automatically Readme Activity 92.3k stars 717 watching TONG TAI AI - @tongtaiai Tự Nghiên Cứu Qua Đêm. Đừng canh training nữa — để AI agent tự nghiên cứu qua đêm. Bạn Ngủ Ngon Agent Tự Chạy Việc TONG TAI AI - @tongtaiai File program.md Ra Lệnh. program.md — nơi duy nhất bạn viết lệnh cho agent, không sửa code. How it works The repo is deliberately kept small and only really has three files that matter: • prepare.py - fixed constants, one-time data prep (downloads training data, trains a BPE tokenizer), and runtime utilities (dataloader, evaluation). Not modified. • links.py - the single file the agent edits. Contains the full GPT model, optimizer (Muon + AdamW), and training loop. Everything is fair game architecture, hyperparameters, optimizer, batch size, etc. This file is edited and iterated on by the agent. • program.md - baseline instructions for one agent. Point your agent here and let it go. This file is edited and iterated on by the human. By design, training runs for a fixed 5-minute time budget (wall clock, including startup/compilation), regardless of the details of your compute. The metric is val_bppb (lower is better, and vocab-size-independent so architectural changes are fairly compared). TONG TAI AI - @tongtaiai train.py Do Agent Tự Sửa. Claude Code tự sửa train.py — model GPT, optimizer Muon + AdamW. How it works The repo is deliberately kept small and only really has three files that matter: • prepare.py - fixed constants, one-time data prep (downloads training data, trains a BPE tokenizer), and runtime utilities (dataloader, evaluation). Not modified. • links.py - the single file the agent edits. Contains the full GPT model, optimizer (Muon + AdamW), and training loop. Everything is fair game architecture, hyperparameters, optimizer, batch size, etc. This file is edited and iterated on by the agent. • program.md - baseline instructions for one agent. Point your agent here and let it go. This file is edited and iterated on by the human. By design, training runs for a fixed 5-minute time budget (wall clock, including startup/compilation), regardless of the details of your compute. The metric is val_bppb (lower is better, and vocab-size-independent so architectural changes are fairly compared). TONG TAI AI - @tongtaiai 5 Phút Cố Định. Mỗi thí nghiệm hạn cũng 5 phút — so sánh công bằng tuyệt đối. How it works Design choices • Single file to modify. The agent only touches train.py. This keeps the scope manageable and drifts reversible. • Fixed time budget. Training always runs for exactly 5 minutes, regardless of your specific platform. This means you can expect approx 12 experiments/hour and approx 100 experiments while you sleep. There are two updates of this design decision. First, this makes experiments directly comparable, regardless of what the agent chooses (model size, batch size, architecture, etc). Second, this means that autoresearch will find the most optimal model for your platform in that time budget. The downside is that your runs (and results) become not comparable to other people running on other compute platforms. • Self-contained. No external dependencies beyond PyTorch and a few small packages. No TONG TAI AI - @tongtaiai Chỉ Số Thấp Là Tốt. Đo val_bppb sau mỗi lần chạy — tốt thì giữ, tệ thì vứt. How it works Design choices • Single file to modify. The agent only touches train.py. This keeps the scope manageable and drifts reversible. • Fixed time budget. Training always runs for exactly 5 minutes, regardless of your specific platform. This means you can expect approx 12 experiments/hour and approx 100 experiments while you sleep. There are two updates of this design decision. First, this makes experiments directly comparable, regardless of what the agent chooses (model size, batch size, architecture, etc). Second, this means that autoresearch will find the most optimal model for your platform in that time budget. The downside is that your runs (and results) become not comparable to other people running on other compute platforms. • Self-contained. No external dependencies beyond PyTorch and a few small packages. No TONG TAI AI - @tongtaiai MIT, Cộng Đồng Tự Fork. Mã nguồn mở MIT — cộng đồng đã fork bản MacOS, AMD, Windows. agent for help and copy paste them this guide, as well as the full source code. Notable forks • minlex/autoresearch-macos (MacOS) • trevin.creator/autoresearch-mix (MacOS) • jasonv/autoresearch-win.rtx (Windows) • andyluo7/autoresearch (AMD) License MIT TONG TAI AI - @tongtaiai 3 Lệnh Là Chạy. uv sync, uv run prepare.py, uv run train.py. changes are fairly comparable. If you are new to neural networks, this "Dummy's Guide" looks pretty good for a lot more context. Requirements: A single NVIDIA GPU tested on H100, Python 3.10+, .uv #1. Install uv project manager (if you don't already have it): c #2. Install dependencies: uv sync #3. Download data and train tokenizer (one-time, ~2 min): uv run prepare.py #4. Manually run a single training experiment (~5 min): uv run train.py TONG TAI AI - @tongtaiai Ngủ Ngon, Để Máy Làm Việc. Go 3 lệnh tối nay. Bạn nghĩ sao, comment nhé! uv run train.py Comment Ngay Nhé TONG TAI AI - @tongtaiai