Neura Market
prompt logo

prompt

Free

Autonomous ML Research Agent for unattended experiment iteration on single-GPU nanochat training.

FreeFree tier
Type
Open Source
Visit Website GitHub (40 stars)

About prompt

A system prompt designed to instruct a large language model to function as an autonomous ML research agent. The agent runs a closed loop of machine-learning experiments on a fixed codebase, modifying train.py, training for a fixed wall-clock budget, measuring a single ground-truth metric (val_bpb, memory_gb), and either keeping or discarding changes. It operates without human intervention, forming hypotheses, running 5-minute experiments, and iterating overnight. Based on the karpathy/autoresearch project (MIT license, 80k+ stars), this prompt enables unattended search through code and hyperparameter space for nano-scale language model training on a single GPU.

Key Features

Closed-loop experiment automation without human intervention
Modifies a single target file (train.py) for architecture/hyperparameter changes
Runs 5-minute experiments with fixed wall-clock budget per trial
Measures ground-truth metrics: validation bits per byte (val_bpb) and memory usage (memory_gb)
Establishes baseline then iterates indefinitely, keeping or discarding changes
Logs results in tab-separated results.tsv with commit, metric, status
Orients by reading current code, results history, and git log before each hypothesis
Forms a single falsifiable experimental idea per iteration

Pros & Cons

Pros
  • Enables fully autonomous, unattended ML research experimentation
  • Structured loop of hypothesis formation, implementation, and validation
  • Reduces human effort for repetitive hyperparameter sweeps
  • Open-source prompt with MIT license, freely usable and modifiable
  • Designed for reproducibility with baseline establishment and versioned changes
Cons
  • Requires a properly set up environment and training data to function
  • Limited to single-GPU nano-scale training scenarios described in the repo
  • May consume significant compute resources over prolonged iteration
  • Not suitable for multi-GPU or large-scale model training experiments
  • Assumes the agent has write access to a specific codebase and Git branch

Best For

Automated hyperparameter optimization for nano-scale language modelsUnattended overnight ML architecture exploration on a single GPUSystematic, reproducible experiment iteration on a fixed codebaseLearning and research in automated machine learning (AutoML) workflows

FAQ

What does the Autonomous ML Research Agent prompt do?
It instructs an LLM to act as an autonomous agent that runs closed-loop ML experiments on a fixed codebase. The agent modifies train.py, trains for a fixed time, evaluates metrics, and iterates without human permission.
What is the source of this prompt?
The prompt is part of the ai-boost/awesome-prompts repository on GitHub and is based on the karpathy/autoresearch project (MIT license, 80k+ stars).